Part-of-speech tagging-based fine-grained information extraction method and system

Abstract

The invention discloses a part-of-speech tagging-based fine-grained information extraction method and system. The method comprises the following steps of: pre-storing codes and information extraction strategies of typical example sentences; phrase-level word segmentation is carried out on a to-be-analyzed statement, and part-of-speech tagging is carried out on phrases; adjacent phrases are merged and hidden according to a coding strategy; according to the part-of-speech and sentence structure marked by S, replacing the nouns, verbs and verb phrases marked as nouns, verbs and verb phrases in the sentences with the part-of-speech to form a code of the to-be-analyzed statement; matching the code of the to-be-analyzed statement with a code of a pre-stored typical example sentence; if the to-be-analyzed sentences are matched with the pre-stored typical example sentences, performing information extraction on the to-be-analyzed sentences according to an extraction strategy of the matched typical example sentences; if not, performing information extraction on the to-be-analyzed sentences according to an information extraction strategy; and the code of the statement to be analyzed and the information extraction strategy are stored. The information extraction process provided by the invention can quickly adapt to new information extraction requirements.

Description

Technical Field

[0001] The present invention relates to the fields of computer data processing and data mining, and in particular to a method for extracting fine-grained information from text based on part-of-speech tagging of an NLP platform. Background Art

[0002] The use and operation of equipment and objects involved in the shipboard support operation are all equipped with corresponding operation guidance documents to guide the implementation of the operation and provide safety tips for operation. The guidance documents contain a wealth of support operation knowledge, but the above knowledge is often distributed in the operation instructions of different equipment and objects to be supported, involving different chapters and clauses, which makes manual retrieval difficult and prone to omissions.

[0003] Knowledge graph technology can integrate isolated data and data associations into a network-structured knowledge system. Based on this network-like knowledge structure system, the effectiveness and completeness of information retrieval results can be effectively improved. The first step in building a knowledge graph is to extract structured data from unstructured or semi-structured text. The technical means to complete this step is called information extraction technology.

[0004] The existing information extraction methods mainly include two methods: rule-based methods and deep learning-based methods. The neural network of the deep learning-based method has problems such as requiring a large amount of corpus annotation and long training time during the training process.

[0005] Standard texts are expressed in unstructured natural language, but they have natural structural characteristics. When extracting information from such texts, rule-based methods can be used. In the actual text processing process, due to differences in writing habits, the characteristics of Chinese expression, and technical limitations of NLP analysis tools, information extraction based solely on rules may result in inaccurate and incomplete decomposition and identification of entities and relationships. Summary of the invention

[0006] The technical problem to be solved by the present invention is the defects existing in the extraction method in the prior art. In order to solve the above problem, a fine-grained information extraction method and system based on part-of-speech tagging are provided.

[0007] The object of the present invention is achieved in the following manner: A fine-grained information extraction method based on part-of-speech tagging, the method comprising the following steps: S1: Get training sentences, use NLP tools to segment sentences at the phrase level, and perform part-of-speech tagging on the decomposed phrases; Adjacent phrases are merged and hidden according to the encoding strategy to form new part-of-speech tags; According to the new part-of-speech tagging and sentence structure, extract the specific part-of-speech sequence, keywords and specific punctuation marks corresponding to the phrases in the sentence as sentence codes; the specific parts of speech include nouns, verbs and verb phrases; the specific punctuation marks are symbols indicating the parallel relationship between words; Extract information from the current sentence by combining the corresponding part of speech of the phrase and the related parts of speech before and after it, characteristic punctuation marks or characteristic words, and save the sentence encoding and corresponding information extraction strategy of the current sentence as the encoding and information extraction strategy of the pre-stored typical example sentence; S2: Segment the sentences to be analyzed at the phrase level and tag the phrases with parts of speech; merge and hide adjacent phrases according to the encoding strategy; S3: According to the parts of speech and sentence structure marked in S2, replace the nouns, verbs and verb phrases in the sentence with parts of speech to form the code of the sentence to be analyzed; S4: matching the encoding of the sentence to be analyzed with the encoding of the pre-stored typical example sentences; S5: If the sentence matches the pre-stored typical example, information is extracted from the sentence to be analyzed according to the extraction strategy of the matched typical example; S6: If there is no match, extract information from the sentence to be analyzed according to the information extraction strategy in S1; and store the encoding of the sentence to be analyzed and the information extraction strategy.

[0008] The information extraction strategy includes: The "subject + predicate + object" sentence structure of NN1 VV NN2 can extract triples: entity 1 represented by NN1 - action relationship embodied by VV - entity 2 represented by NN2, entities and relationships; The attributive phrase of NN1 DEG NN2 can extract triples: entity 2 represented by NN2—belonging relationship—entity 1 represented by NN1; The attributive phrase of AD DEG NN can extract the attribute represented by AD in the entity represented by NN; The sentence structure of "subject + predicate + parallel object" of NN1 VV NN2...NNn can extract triples: entity 1 represented by NN1 - action relationship embodied by VV - entity 2 represented by NN2, ..., entity 1 represented by NN1 - action relationship embodied by VV - entity n represented by NNn; "Subject + predicate + object" or "predicate + object", sentence ① + "before" | "after" + "subject + predicate + object" or "predicate + object", sentence ②, can extract the process triplet: operation process ① represented by sentence ① - the sequence relationship of operation processes - operation process ② represented by sentence ②.

[0009] The encoding strategy includes: A. For consecutive nouns, mark their part of speech as noun (NN), noun phrase (NP), person name (NR), or proper noun (NT), and combine them to generate a new noun (NN). B. Combine phrases with the part of speech marked as adjective (JJ) + noun (NN), and the part of speech of the combined phrase is noun (NN). C. Combine consecutive verbs with the part of speech marked as verb (VV) or verb phrase (VP) to generate a new verb. D. Combine phrases with the part of speech marked as adverb (AD) + verb (VV), and the part of speech of the combined phrase is verb (VV). E. Combine phrases with the part of speech marked as verb (VV) + preposition (P), and the part of speech of the combined phrase is verb phrase (VP). F. When there is a combination of noun (NN1) + '的' (DEG) + noun (NN2), hide NN1 and DEG, and only retain noun NN2. G. When there is a combination of adverb (AD) + '的' (DEG) + noun (NN), hide AD and DEG, and only retain noun NN. H. When there is a combination of noun (NN1) + locative word (LC) + '的' (DEG) + noun (NN2), hide NN1, LC, and DEG, and only retain noun NN2. I. Hide the determiners in the sentence, where the determiners include '全部' (all), '所有' (all), and '其它' (other). J. Hide all the content from '等' (ETC) to the end of the sentence. K. Hide the segmentation symbols such as '《' and '》'. After combining and hiding the phrases in the sentence according to the above strategy, replace the nouns with the part of speech of NN, the verbs with the part of speech of VV, and the verb phrases with the part of speech of VP. The generated symbol string sequence is the sentence encoding.

[0010] In step S4, the encoding of the sentence to be analyzed and the encoding of the pre - stored typical example sentences are pattern - matched using a method based on the similarity of sentence structures. Specifically: the differences between sentences are generally represented by the edit distance. In the actual sentence pattern matching process, first check whether the number of verbs in the sentence to be matched is the same as that in the sentence pattern. The sentence to be matched only compares the edit distance with the sentence encodings with the same number of verbs, and then finds the sentence pattern with the smallest edit distance from its sentence encoding as the information extraction template.

[0011] The information extraction includes extracting physical entities, ownership relationships, action relationships, operation process entities, and the sequence of operation processes.

[0012] A system for implementing the above - mentioned fine - grained information extraction method based on part - of - speech tagging, the system includes: The module for editing and saving the text to be analyzed is used to browse and modify the text to be analyzed; Sentence analysis and part-of-speech tagging module: Use NLP tools to segment sentences and tag the decomposed phrases with parts of speech; Relational entity extraction and preservation module: uses the results of sentence analysis and part-of-speech tagging to identify entities in sentences and the relationships between entities; Custom phrase editing module: It can identify possible proprietary words based on the results of relational entity extraction, and can edit them through the human-computer interaction interface and save them to the specific files of the NLP tool; Entity and relationship editing module: provides an editing interface for editing and correcting the relationships, entities, and related label attributes identified from sentences.

[0013] Beneficial effects of the present invention: Compared with the extraction method using deep learning technology, the present invention does not require tedious data labeling and model tuning; at the same time, compared with the traditional extraction method using syntactic analysis, the implementation of the extraction process does not require special induction of information extraction rules, and has strong scalability, clear operating mechanism, and strong interpretability of the extraction results.

[0014] At the same time, its phrase merging based on parts of speech helps to identify new entities in the information extraction process.

[0015] In addition, this method involves sentence classification and knowledge extraction mode functions, which are all customized by the user in the early stage. Therefore, the information extraction process can quickly adapt to new information extraction needs and can be well migrated and applied to other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a schematic diagram of the principle of the knowledge extraction method based on sentence structure matching.

[0017] Figure 2 It is a schematic diagram of sentence encoding based on part-of-speech tagging.

[0018] Figure 3 It is a sentence encoding comparison based on part-of-speech tagging.

[0019] Figure 4 This is a schematic diagram of the functional relationship of the automatic extraction tool software for ship surface support operation information.

[0020] Figure 5 It is a graphical display of the results of sentence information extraction containing multiple verbs and parallel objects.

[0021] Figure 6 It is a graphical display of the sentence information extraction results including the process. DETAILED DESCRIPTION

[0022] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0023] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same technical meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0024] A fine-grained information extraction method based on part-of-speech tagging, the method comprising the following steps: S1: Get training sentences, use NLP tools to segment sentences at the phrase level, and perform part-of-speech tagging on the decomposed phrases; Adjacent phrases are merged and hidden according to the encoding strategy to form new part-of-speech tags; According to the new part-of-speech tagging and sentence structure, extract the specific part-of-speech sequence, keywords and specific punctuation marks corresponding to the phrases in the sentence as sentence codes; the specific parts of speech include nouns, verbs and verb phrases; the specific punctuation marks are symbols such as semicolons that indicate parallel relationships between words; Extract information from the current sentence by combining the corresponding part of speech of the phrase and the related parts of speech before and after it, characteristic punctuation marks or characteristic words, and save the sentence encoding and corresponding information extraction strategy of the current sentence as the encoding and information extraction strategy of the pre-stored typical example sentence; S2: Segment the sentences to be analyzed at the phrase level and tag the phrases with parts of speech; merge and hide adjacent phrases according to the encoding strategy; S3: According to the parts of speech and sentence structure marked in S2, replace the nouns, verbs and verb phrases in the sentence with parts of speech to form the code of the sentence to be analyzed; S4: matching the encoding of the sentence to be analyzed with the encoding of the pre-stored typical example sentences; S5: If the sentence matches the pre-stored typical example, information is extracted from the sentence to be analyzed according to the extraction strategy of the matched typical example; S6: If there is no match, extract information from the sentence to be analyzed according to the information extraction strategy in S1; and store the encoding of the sentence to be analyzed and the information extraction strategy.

[0025] The information extraction strategy includes: The "subject + predicate + object" sentence structure of NN1 VV NN2 can extract triples: entity 1 (represented by NN1) - action relationship (represented by VV) - entity 2 (represented by NN2), entities and relationships; A modifier-head phrase of "NN1 DEG NN2" can extract a triple: Entity 2 (represented by NN2) - belonging relationship - Entity 1 (represented by NN1); A modifier-head phrase of "AD DEG NN" can extract the attribute (represented by AD) in the entity (represented by NN); A sentence pattern of "subject + predicate + coordinated objects" of "NN1 VV NN2…NNn" can extract triples: Entity 1 represented by NN1 - action relationship reflected by VV - Entity 2 represented by NN2,..., Entity 1 represented by NN1 - action relationship reflected by VV - Entity n represented by NNn; "Subject + predicate + object" or "predicate + object", Sentence ① + "before" | "after" + "subject + predicate + object" or "predicate + object", Sentence ②, can extract a process triple: Operation process ① represented by Sentence ① - operation process sequence relationship - Operation process ② represented by Sentence ②.

[0026] The described encoding strategy includes: A. Merge continuously occurring nouns (with part-of-speech tags NN (noun), NP (noun phrase), NR (person name), NT (proper noun. institutional group name)) to generate a new noun; B. Merge phrases with part-of-speech tags JJ+NN (adjective + noun), and the part-of-speech of the merged phrase is NN (noun); C. Merge continuously occurring verbs (with part-of-speech tags VV, VP) to generate a new verb; D. Merge phrases with part-of-speech tags AD+VV (adverb + verb), and the part-of-speech of the merged phrase is VV (verb); E. Merge phrases with part-of-speech tags VV+P (verb + preposition), and the part-of-speech of the merged phrase is VP (verb phrase); F. When a combination of noun NN1 + 'of' DEG + noun NN2 appears, hide NN1 and DEG, and only retain noun NN2; G. When a combination of adverb AD + 'of' DEG + noun NN appears, hide AD and DEG, and only retain noun NN; H. When a combination of noun NN1 + locative LC + 'of' DEG + noun NN2 appears, hide NN1, LC, and DEG, and only retain noun NN2; I. Hide the determiners in the sentence, and the determiners include 'all', 'every', 'other'; J. Hide all the content from 'etc.' ETC to the end of the sentence; K. Hide the segmentation symbols such as '《' and '》'; After merging and hiding the phrases in the sentence according to the above strategy, nouns with part of speech NN, verbs with part of speech VV and verb phrases with part of speech VP are replaced with part of speech, and the generated symbol string sequence is the sentence encoding.

[0027] In S5, the encoding of the sentence to be analyzed is matched with the encoding of the pre-stored typical example sentence using a method based on sentence structure similarity. Specifically, the difference between sentences is generally represented by the edit distance. In the actual sentence pattern matching process, it is first checked whether the number of verbs contained in the sentence to be matched is the same as that contained in the sentence pattern. The sentence to be matched is only compared with the sentence encoding with the same number of verbs for the edit distance, and then the sentence pattern with the smallest edit distance with its sentence encoding is found as the information extraction template.

[0028] The information extraction includes extracting physical entities, belonging relationships, action relationships, operation process entities, operation process sequence relationships, etc.

[0029] Sentence encoding: The following is an example of the requirements for mooring ropes: "The mooring ropes shall be connected from the specific mooring accessories of the aircraft to the deck mooring accessories (hereinafter referred to as sentence ①), and shall not press on other aircraft parts such as hydraulic shock absorbers, hydraulic pipes, tires, etc. (hereinafter referred to as sentence ②)" to illustrate the information extraction process: After treating "mooring rope" and "mooring accessories" as proper nouns, the sentences are segmented and tagged with parts of speech, and the adjacent phrases are merged and hidden according to the above merging rules: "should (VV)" and "from (P)" are merged into "should from (VV)"; "specific (JJ)" and "mooring accessories (NN)" are merged into "specific mooring accessories (NN)"; "connect (VV)" and "to (VV)" are merged into "connected to (VV)", and "deck (NN)" and "mooring accessories (NN)" are merged into "deck mooring accessories (NN)". Figure 2 shown.

[0030] Sentence ① generates the analysis result of {subject (mooring rope, NN)-predicate (should follow, VV)-{object {subordinate structure (aircraft-mooring accessories, NN)}-predicate (connection, VV)-object (deck mooring accessories, NN)}} according to the marked part of speech and the above-mentioned information extraction strategy; After completing the subject according to the context of sentence ②, the analysis result of {subject (mooring rope, NN)-predicate (no oppression, VV)-object {parallel structure {phrase 1 (hydraulic shock absorber, NN), phrase 2 (hydraulic pipe, NN), phrase 2 (tire, NN)}}} is generated according to the definition of part of speech and sentence structure.

[0031] According to the above sentence encoding strategy, the sentence encoding (template) generated by this sentence is: ①: NN VV NN VV NN; ②: VVNN, NN, NN.

[0032] Sentence matching: Now, the sentence "Qualified personnel shall perform pre-check operations in accordance with the provisions of the Naval Aviation Training and Operations Procedure Standardization Manual." (hereinafter referred to as sentence ③) is encoded according to the above strategy and compared with the above-mentioned sentence ①.

[0033] After pre-processing "Naval Aviation Training and Operational Procedure Standardization Manual", "Pre-check Operations", "Qualified" and other words as proprietary words and special phrases, phrases and parts of speech are merged and hidden according to "Rule 1" to "Rule 8". The results are compared with the analysis results of sentence ①. Figure 3 shown.

[0034] The feature vectors of both sentences ① and ③ are: NN VP NN VV NN, and their edit distance is 0. Therefore, the information extraction method of sentence ③ can be completely carried out in the same way as sentence ①, without the need to extract information through syntactic analysis again.

[0035] The sentence analysis function of the ship surface support operation information extraction tool compiled according to the above method is implemented using the TexSmart offline SDK provided by Tencent Artificial Intelligence Laboratory. Since there are many professional terms and special words in this knowledge field, the effect of using the general vocabulary of the pre-trained language model alone to segment and tag the text is not ideal. The technical characteristics of TexSmart offline SDK can add custom words and entities, sort out the proprietary words and special phrases in this knowledge field, and add them to the predefined file corresponding to TexSmart offline SDK in the specified format. In the subsequent analysis, the SDK will understand the above words as a whole and no longer decompose and analyze them, ensuring the rationality of the word segmentation results in this field. At the same time, the information extraction tool also has the function of generating custom words from entity names, and can dynamically supplement and adjust custom words according to the information extraction results. The content to be analyzed and the analysis results involved in the information extraction tool are saved locally in the form of xml files, which is convenient for comparing and checking the extraction results.

[0036] The logical relationship between the various functions of the tool software is as follows: Figure 4 shown.

[0037] Functions of tool-related modules: Text editing and saving module for analysis, used for browsing and modifying text to be analyzed; sentence analysis and part-of-speech tagging module, which uses TexSmart offline SDK to segment sentences and tag the decomposed phrases with parts of speech; relational entity extraction and saving module, which uses the results of sentence analysis and part-of-speech tagging to identify entities in sentences and the relationship between entities; custom phrase editing module, which can identify possible proprietary vocabulary based on the results of relational entity extraction, and can edit and save it to a specific file of TexSmart offline SDK through the human-computer interaction interface; entity and relationship editing module, which provides an editing interface for editing and correcting the relationships, entities and related tag attributes identified from the sentences; Cypher statement generation module, which generates Cypher statements for the currently extracted entities and relationships; xml file access and parsing, which provides the data xml stream access and parsing functions required by the software.

[0038] Information Extraction Example For the above sentences ① and ②, according to the knowledge model of ship surface support operations and combined with the analysis of the above sentence structures, the information extraction results are as follows: Physical entities (7): mooring ropes, aircraft, mooring accessories, deck mooring accessories, hydraulic shock absorbers, hydraulic pipes, tires, etc.; Affiliation (1): Aircraft - tie-down accessories; Action relationships (5): Mooring rope - <should> be connected to the mooring accessories - deck mooring accessories, Mooring rope - <must not> press on the hydraulic shock absorber, Mooring rope - <must not> press on the hydraulic pipe, Mooring rope - <must not> press on the tire.

[0039] Based on the above extraction results, generate a Cypher statement and submit it to Neo4j for interpretation and execution. The graphical display results are as follows: Figure 5 shown.

[0040] The analysis process of clauses with processes is similar to the above, except that the sentences to be analyzed are first divided into different process clauses by keywords representing the sequence of processes, and entities representing process activities need to be created when entities are created later. For example, "Before the aircraft is separated from the tow bar, the parking brakes of the aircraft and the tow vehicle are lowered, and the wheel chocks are installed." (hereinafter referred to as sentence ④) Due to the presence of the keyword "before" in the sentence, the relevant sentence part is decomposed into three processes: "the aircraft is separated from the tow bar", "the parking brakes of the aircraft and the tow vehicle are lowered", and "the wheel chocks are installed".

[0041] Similarly, based on the knowledge model of ship surface support operations and combined with the analysis of the above sentence structure, the information extraction results are as follows: Physical entities (5): parking brake, tractor, aircraft, drawbar, wheel chock; Operation process (3): disengage the drawbar, lower the parking brake, and install the wheel chock; Affiliation (2): tractor-parking brake, aircraft-parking brake; Action relationship (1): aircraft-installation-wheel chock; The process involves 3 entity relationships: drawbar-disengage-disengage the drawbar, parking brake-lower-lower the parking brake, wheel chock-install-install wheel chock; The sequence of processes (2): before disengaging the tow bar, lower the parking brake; before disengaging the tow bar, install the wheel chocks.

[0042] Based on the above extraction results, generate a Cypher statement and submit it to Neo4j for interpretation and execution. The graphical display results are as follows: Figure 6 shown.

[0043] The above is only a preferred embodiment of the present invention. It should be pointed out that for those skilled in the art, several changes and improvements can be made without departing from the overall concept of the present invention, which should also be regarded as the scope of protection of the present invention.

Claims

1. A fine-grained information extraction method based on part-of-speech tagging, characterized by: The method comprises the following steps: S1: Get training sentences, use NLP tools to segment sentences at the phrase level, and perform part-of-speech tagging on the decomposed phrases; Adjacent phrases are merged and hidden according to the encoding strategy to form new part-of-speech tags; According to the new part-of-speech tagging and sentence structure, extract the specific part-of-speech sequence, keywords and specific punctuation marks corresponding to the phrases in the sentence as sentence codes; the specific parts of speech include nouns, verbs and verb phrases; the specific punctuation marks are symbols indicating the parallel relationship between words; Extract information from the current sentence by combining the corresponding part of speech of the phrase and the related parts of speech before and after it, characteristic punctuation marks or characteristic words, and save the sentence encoding and corresponding information extraction strategy of the current sentence as the encoding and information extraction strategy of the pre-stored typical example sentence; S2: Segment the sentences to be analyzed at the phrase level and tag the phrases with parts of speech; merge and hide adjacent phrases according to the encoding strategy; S3: According to the parts of speech and sentence structure marked in S2, replace the nouns, verbs and verb phrases in the sentence with parts of speech to form the code of the sentence to be analyzed; S4: matching the encoding of the sentence to be analyzed with the encoding of the pre-stored typical example sentences; S5: If the sentence matches the pre-stored typical example, information is extracted from the sentence to be analyzed according to the extraction strategy of the matched typical example; S6: If there is no match, extract information from the sentence to be analyzed according to the information extraction strategy in S1; and store the encoding of the sentence to be analyzed and the information extraction strategy.

2. The fine-grained information extraction method based on part-of-speech tagging according to claim 1 is characterized in that: The information extraction strategy includes: The "subject + predicate + object" sentence structure of NN1 VV NN2 can extract triples: entity 1 represented by NN1 - action relationship embodied by VV - entity 2 represented by NN2, entities and relationships; The attributive phrase of NN1 DEG NN2 can extract triples: entity 2 represented by NN2—belonging relationship—entity 1 represented by NN1; The attributive phrase of AD DEG NN can extract the attribute represented by AD in the entity represented by NN; The sentence structure of "subject + predicate + parallel object" of NN1 VV NN2...NNn can extract triples: entity 1 represented by NN1 - action relationship embodied by VV - entity 2 represented by NN2, ..., entity 1 represented by NN1 - action relationship embodied by VV - entity n represented by NNn; "Subject + predicate + object" or "predicate + object", sentence ① + "before" | "after" + "subject + predicate + object" or "predicate + object", sentence ②, the process triplet can be extracted: the operation process ① represented by sentence ① - the sequence relationship of the operation processes - the operation process ② represented by sentence ②.

3. The fine-grained information extraction method based on part-of-speech tagging according to claim 1 is characterized in that: The encoding strategy includes: A. Mark the nouns that appear consecutively as noun NN, noun phrase NP, personal name NR, and proper noun NT, and merge them to generate a new noun NN; B. Combine the phrases with the part-of-speech tags of adjective JJ + noun NN, and the part-of-speech of the combined phrase is noun NN; C. Combine the continuously occurring verbs with the part-of-speech tags of VV and VP to generate a new verb VV; D. Combine the phrases with the part-of-speech tags of adverb AD + verb VV, and the part-of-speech of the combined phrase is verb VV; E. Combine the phrases with the part-of-speech tags of verb VV + preposition P, and the part-of-speech of the combined phrase is verb phrase VP; F. When the combination of noun NN1 + '的' DEG + noun NN2 appears, hide NN1 and DEG, and only retain noun NN2; G. When the combination of adverb AD + '的' DEG + noun NN appears, hide AD and DEG, and only retain noun NN; H. When the combination of noun NN1 + locative LC + '的' DEG + noun NN2 appears, hide NN1, LC, and DEG, and only retain noun NN2; I. Hide the determiners in the sentence, and the determiners include '全部' (all), '所有' (all), '其它' (other); J. Hide all the content from '等' ETC to the end of the sentence; K. Hide the segmentation symbols such as '《' and '》'; After combining and hiding the phrases in the sentence according to the above strategy, replace the nouns with the part-of-speech of NN, the verbs with the part-of-speech of VV, and the verb phrases with the part-of-speech of VP with their part-of-speech, and the generated symbol string sequence is the sentence encoding.

4. The fine-grained information extraction method based on part-of-speech tagging according to claim 1 is characterized in that: In S4, the encoding of the sentence to be analyzed and the encoding of the pre-stored typical example sentences are pattern-matched using a method based on the sentence structure similarity. Specifically: the differences between sentences are generally represented by the edit distance. In the actual sentence pattern matching process, first check whether the number of verbs contained in the sentence to be matched is the same as that in the sentence pattern. The sentence to be matched only compares the edit distance with the sentence encoding of the sentence with the same number of verbs, and then finds the sentence pattern with the smallest edit distance from its sentence encoding as the information extraction template.

5. The fine-grained information extraction method based on part-of-speech tagging according to claim 1 is characterized in that: The information extraction includes extracting physical entities, ownership relationships, action relationships, operation process entities, and the sequence of operation processes.

6. A system for implementing the fine-grained information extraction method based on part-of-speech tagging according to any one of claims 1 to 5, characterized in that: The system includes: Module for editing and saving the text to be analyzed: used for browsing and modifying the text to be analyzed; Module for sentence analysis and part-of-speech tagging: uses NPL tools to segment the sentence and tag the part-of-speech of the decomposed phrases; Module for extracting and saving relationship entities: uses the results of sentence analysis and part-of-speech tagging to identify the entities in the sentence and the association relationships between the entities; Custom phrase editing module: can identify possible proper nouns according to the results of relationship entity extraction, and can be edited through the human-computer interaction interface and saved to a specific file of the NPL tool; Entity and relationship editing module: provides an editing interface to edit and correct the relationships, entities, and related tag attributes identified from the sentence.

Images