A professional word segmentation method applied to nuclear power industry

Abstract

The application belongs to the field of natural language processing in the nuclear power industry, and particularly relates to a professional word segmentation method applied to nuclear power industry corpus, which comprises nuclear power professional vocabulary construction, nuclear power stop word vocabulary construction, nuclear power synonym vocabulary construction, nuclear power same pronoun vocabulary construction, nuclear power field new word recognition, nuclear power field entity automatic recognition, nuclear power field synonym automatic recognition, nuclear power corpus Chinese accurate word segmentation and the like operations. The application has the beneficial effect of completely solving the problem of inaccurate Chinese word segmentation of nuclear power industry corpus, laying a solid foundation for the application of subsequent big data and artificial intelligence in the field of nuclear power natural language processing, and reducing the investment of other practitioners in the field of nuclear power in natural language processing.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing in the nuclear power industry, specifically involving a professional word segmentation method for nuclear power industry corpora. It can realize functions such as Chinese word segmentation, stop word processing, automatic new word recognition, automatic entity recognition, and automatic synonym recognition for nuclear power corpora, thus meeting the requirements for accurate word segmentation of nuclear power corpora. Background Technology

[0002] With the application of technologies such as big data, machine learning, knowledge graphs, search engines, and intelligent question answering in various fields of the nuclear power industry, research on text processing (natural language processing) in the nuclear power industry is increasing. From the perspective of the development of artificial intelligence, natural language processing is considered relatively difficult, especially for Chinese text processing. Natural language processing aims to achieve natural language communication between humans and machines, or to achieve natural language understanding and generation. The fundamental reason for the difficulty of natural language processing is the widespread ambiguity or polysemy at various levels of natural language text and dialogue. Unlike English, there are no spaces between words in Chinese sentences; therefore, Chinese word segmentation is considered a fundamental step in Chinese natural language processing. Inaccurate word segmentation makes feature engineering difficult, affecting the machine's understanding of semantics. The accuracy of semantic search, intelligent question answering, machine learning, and other functions, as well as the user experience, will be significantly impacted.

[0003] The main difficulties in Chinese word segmentation are as follows:

[0004] 1. Lexicon-based word segmentation methods often cannot identify new words or domain-specific terms;

[0005] 2. The ambiguity of word segmentation cannot be resolved by machines. Chinese text, in form, is a string composed of Chinese characters (including punctuation marks). Characters form words, words form phrases, phrases form sentences, and sentences form paragraphs, sections, chapters, and articles. Ambiguity and polysemy exist in characters, words, phrases, sentences, paragraphs, sections, chapters, and articles; that is, strings that appear identical in form can be interpreted as different word strings, phrase strings, etc., and have different meanings in different scenarios or contexts. Resolving ambiguity requires an extremely large amount of knowledge and reasoning. How to collect and organize this knowledge comprehensively; how to find a suitable format to store it in a computer system; and how to effectively use it to eliminate ambiguity are all extremely laborious and difficult tasks.

[0006] For the nuclear power industry, Chinese word segmentation of nuclear power-related corpora faces the same problems, and there is currently no professional word segmentation tool suitable for nuclear power-related corpora. Open-source word segmentation tools are extremely inaccurate when segmenting nuclear power industry corpora. To solve this problem, it is necessary to train the word segmentation model with a large-scale corpus of real nuclear power texts, and to compile a large-scale, information-rich nuclear power-related lexicon.

[0007] The present invention aims to solve the above problems by constructing a professional word segmentation tool for the nuclear power industry, which can realize functions such as Chinese word segmentation, stop word processing, automatic new word recognition, automatic entity recognition, and automatic synonym recognition for nuclear power corpus, thereby meeting the requirements for accurate word segmentation of nuclear power corpus. Summary of the Invention

[0008] In response to the current situation in the field of natural language processing in the nuclear power industry, such as inaccurate Chinese word segmentation, lack of identification of new words in specific domains, and inadequate semantic disambiguation in specific domains, the purpose of this invention is to provide a professional word segmentation method for the nuclear power industry. Taking nuclear power corpus generated in various business processes of nuclear power plants as the research object, this invention realizes functions such as Chinese word segmentation, stop word processing, automatic identification of new words, automatic identification of entities, and automatic identification of synonyms for nuclear power corpus, thereby meeting the requirements for accurate word segmentation of nuclear power corpus.

[0009] The technical solution of this invention is as follows: A professional word segmentation method applied to the nuclear power industry, comprising the following steps:

[0010] Step 1: Establish a nuclear power-related thesaurus by analyzing the corpus;

[0011] Step 2: Establish a nuclear power plant delimited word database;

[0012] Step 3: Establish a thesaurus of nuclear power synonyms;

[0013] Step 4: Establish a unified thesaurus of pronouns for nuclear power;

[0014] Step 5: Combine the nuclear power professional terminology database, nuclear power stop word database, nuclear power synonym database, nuclear power common pronoun database, and jieba general terminology database constructed in steps 1 to 4 above to form a nuclear power terminology database;

[0015] Step 6: By employing an encoder-decoder model to perform machine learning on the nuclear power industry corpus, automatic new word recognition can be achieved on the nuclear power industry corpus;

[0016] Step 7: Supervised machine learning is performed on the nuclear power industry corpus using an encoder-decoder model. By annotating and learning the nuclear power entity categories from a portion of the nuclear power industry corpus, automatic entity recognition of the nuclear power industry corpus is achieved.

[0017] Step 8: Cluster the “NPP-D” lexicon constructed from the above corpus, and then use the latent semantic retrieval model and singular value decomposition algorithm on each cluster to achieve automatic synonym recognition of nuclear power industry corpus in an unsupervised learning manner. The identified synonyms are included in the nuclear power synonym lexicon after being reviewed by professionals.

[0018] Step 9: Cluster the “NPP-D” lexicon constructed from the above corpus, and then use the latent semantic retrieval model and singular value decomposition algorithm on each cluster to automatically identify the same pronouns in the nuclear power industry corpus in an unsupervised learning manner. The identified same pronouns are included in the nuclear power same pronoun lexicon after being reviewed by professionals.

[0019] Step 10: Based on the “NPP-D” lexicon, a nuclear power word segmentation model is constructed using a dynamic programming algorithm to find the maximum probability path and a hidden Markov model in an unsupervised learning manner;

[0020] Step 11: Based on the nuclear power word segmentation model, establish a nuclear power word segmentation tool. The nuclear power word segmentation tool achieves accurate Chinese word segmentation of nuclear power corpus by calling the nuclear power word segmentation model in Step 10.

[0021] Step 1 involves compiling various types of nuclear power terminology by searching through system design manuals, equipment delivery manuals, structure layout design documents, equipment failure mode documents, as well as equipment lists, material lists, organizational lists, personnel information lists, and risk lists in the nuclear power plant's production management system.

[0022] Step 2, the construction of the stop words, includes the following aspects:

[0023] List all punctuation marks as stop words;

[0024] All special characters except "#", uppercase and lowercase letters and numbers are listed as stop words;

[0025] List all modal particles as stop words;

[0026] The stop word dictionary consists of fields such as "stop word name" and "creation date".

[0027] In step 4, the term "same pronoun lexicon" refers to two words that, when used in the nuclear power industry, refer to the same thing. The process of constructing the same pronoun lexicon is as follows:

[0028] The device code and device name of the same device are grouped into the same pronoun and imported into the same pronoun dictionary;

[0029] The system number and system name of the same system are grouped into the same pronoun and imported into the same pronoun dictionary;

[0030] The factory codes and factory names of the same factory building are grouped into the same pronoun and imported into the same pronoun dictionary;

[0031] Group the room codes and room names of the same room into the same pronoun and import them into the same pronoun dictionary;

[0032] The Chinese and English names of nuclear power plant manufacturers from the same manufacturer are grouped into the same pronoun and imported into the same pronoun dictionary.

[0033] The Chinese and English names of nuclear power plant contractors under the same contractor are grouped into the same pronoun and imported into the same pronoun dictionary.

[0034] The material codes and material names of the same material are grouped into the same pronoun and imported into the same pronoun thesaurus.

[0035] The same pronoun dictionary consists of the fields of "synonym list", "creation date" and "association relationship".

[0036] In step 6, a dictionary lookup tree algorithm is used to automatically identify the frequency of new words, and a hidden Markov model is used to automatically identify the part of speech of new words. The identified new words are automatically included in the nuclear power professional thesaurus. The specific process of new word identification is as follows:

[0037] The encoder model utilizes a Long Short-Term Memory (LSTM) neural network to transform each character of nuclear power-related technical terms into a 768-dimensional feature vector using the BERT framework. These vectors are concatenated to form a feature matrix corresponding to the sentence, which is then fed into the LTM neural network. The LTM network calculates a matrix containing the boundary features of the nuclear power-related technical terms, which is then passed to the decoder. The decoder uses a Conditional Random Field (CRF) to calculate the probability of each character representing a boundary, completing word segmentation. Within the segmented sentences, a Hidden Markov Model (HMM) is employed to calculate the part-of-speech tag (POS) of each technical term using statistical conditional probability. After new word recognition, a dictionary lookup tree is constructed for all new words, and these are matched against all documents. The number of matches performed for each word is taken as its frequency.

[0038] The automatic entity recognition process in step 7 is as follows:

[0039] The encoder model utilizes a long short-term memory neural network to transform each character of nuclear power-related technical terms into a 768-dimensional feature vector using the BERT framework. The vectors are concatenated to form a feature matrix corresponding to the sentence. This matrix is ​​then fed into the long short-term memory neural network, which calculates a matrix containing the features of the nuclear power-related technical terms. This feature matrix is ​​then fed into the decoder, where a conditional random field is used to calculate the probability of each character as a boundary, thus completing word segmentation and entity label recognition.

[0040] The automatic synonym recognition process in step 8 is as follows:

[0041] For each word in the NPP-D vocabulary, a semantic vector is calculated using a word embedding model. The vectors are then clustered into several clusters using the k-Means clustering algorithm. Within each cluster, the word feature vectors are concatenated to construct a feature matrix. The feature matrix is ​​then decomposed into three different matrices: a left matrix, a singular value diagonal matrix, and a right matrix, using a singular value decomposition algorithm. The values ​​in the singular value diagonal matrix are analyzed, and singular values ​​exceeding a certain threshold are filtered out. The corresponding columns in the left matrix are then searched, and the word pairs with the closest similarity are selected as synonyms.

[0042] The automatic identification process for the same pronoun in step 8 is as follows:

[0043] For each word in the NPP-D vocabulary, a semantic vector is calculated using a word embedding model. The vectors are then clustered into several clusters using the k-Means clustering algorithm. Within each cluster, the word feature vectors are concatenated to construct a feature matrix. The feature matrix is ​​then decomposed into three different matrices: a left matrix, a singular value diagonal matrix, and a right matrix using a singular value decomposition algorithm. The right matrix after decomposition is the semantic similarity matrix. Word pairs with high semantic similarity are set as the same pronoun.

[0044] In step 10, the hidden Markov model marks the frequency of each word in the existing nuclear power corpus as its probability of occurrence, sets the conditional probability of new words to the minimum frequency, and trains the existing nuclear power professional vocabulary in an unsupervised manner to obtain the probability matrix of the sentences. The length of the matrix corresponds to the length of the sentences, and the width corresponds to the label category. Through a dynamic programming pathfinding algorithm, a path with the highest comprehensive probability is calculated in the probability matrix. Based on the matrix row and column numbers traversed by the path, BMES (Begin-Middle-End-Single) labeling is performed to mark the boundary of each word, and the corresponding label is used as each character in the sentence.

[0045] The beneficial effects of this invention are as follows: This invention constructs a professional word segmentation method applicable to the nuclear power industry, which completely solves the problem of inaccurate Chinese word segmentation in nuclear power industry corpora, lays a solid foundation for the subsequent application of big data and artificial intelligence in the field of nuclear power natural language processing, and reduces the investment of other practitioners in the nuclear power field in natural language processing. Attached Figure Description

[0046] Figure 1 The flowchart of a professional word segmentation tool for the nuclear power industry provided by the present invention. Detailed Implementation

[0047] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] This invention constructs a professional word segmentation tool suitable for the nuclear power industry, including functions such as building a nuclear power professional lexicon, building a nuclear power stop word lexicon, building a nuclear power synonym lexicon, building a nuclear power common pronoun lexicon, identifying new words in the nuclear power field, automatically identifying entities in the nuclear power field, automatically identifying synonyms in the nuclear power field, and accurately segmenting Chinese words from nuclear power corpus.

[0049] This invention mainly involves the following aspects:

[0050] (1) This invention establishes a nuclear power professional lexicon by sorting through the corpus of nuclear power industry;

[0051] (2) This invention establishes a nuclear power stop word lexicon by sorting through the corpus of nuclear power industry;

[0052] (3) This invention establishes a thesaurus of nuclear power synonyms by sorting through the corpus of nuclear power industry;

[0053] (4) This invention establishes a lexicon of nuclear power-related pronouns by sorting through the corpus of nuclear power industry;

[0054] (5) This invention uses machine learning to train nuclear power industry corpus, realizes automatic new word recognition of nuclear power industry corpus, automatically identifies word frequency and part of speech in new words, and automatically incorporates the identified new words into nuclear power professional lexicon;

[0055] (6) This invention uses machine learning to train nuclear power industry corpus, thereby achieving automatic entity recognition of nuclear power industry corpus and automatically classifying the new words identified in step (5).

[0056] (7) This invention achieves automatic synonym recognition of nuclear power industry corpus by training the corpus with machine learning;

[0057] (8) This invention achieves automatic identification of the same pronoun in nuclear power industry corpus by training the corpus with machine learning;

[0058] (9) Based on (1) to (8), this invention constructs a nuclear power word segmentation model;

[0059] (10) Based on the nuclear power word segmentation model, this invention develops a nuclear power word segmentation interface and forms a nuclear power word segmentation tool;

[0060] (11) The nuclear power corpus to be segmented is accurately segmented using the nuclear power segmentation tool.

[0061] The nuclear power industry corpus involved in this invention includes, but is not limited to, status reports, operational event reports, internal event reports, corrective action plans, WANO event reports, IAEA event reports, timely experience feedback, work orders, work order completion reports, quality defect reports, non-conformity reports, change requests, change management, equipment data, material master data, periodic testing, operational procedures, isolation data, work permit control data, work permit commencement data, radiation protection permits, general permits, daily planning and scheduling data, overhaul planning and scheduling data, nuclear power plant outlines, nuclear power plant regulations, and nuclear power plant manuals.

[0062] like Figure 1 As shown, a professional word segmentation method applied to the nuclear power industry includes the following steps:

[0063] Step 1: Establish a nuclear power professional thesaurus by sorting through the corpus.

[0064] This involves searching through system design manuals, equipment delivery manuals, structure layout design documents, equipment failure mode documents, and equipment lists, material lists, organizational structures, personnel information lists, and risk lists from nuclear power plant production management systems to compile various types of nuclear power-related terminology databases. During this process, duplicate terminology databases from similar power plants may be encountered. In such cases, deduplication is necessary. The nuclear power terminology database consists of fields such as "words," "frequency," "part of speech," "custom attributes," "data source," "project to which it belongs," "creation date," and "update date."

[0065] The nuclear power terminology database includes categories such as equipment codes, equipment names, system numbers, system names, plant codes, plant names, room codes, room names, Chinese and English names of nuclear power plant manufacturers, Chinese and English names of nuclear power plant contractors, risk keywords, failure modes, fault modes, causes of failure, material codes, material names, personnel categories, personnel names, organization names, reactor or plant status, location numbers, location names, plant activities, event consequences, work group names, and nuclear power plant department names for all nuclear power plants in China. These categories are also known as nuclear power entity categories, and each nuclear power entity category includes multiple terms. Each term is assigned a part-of-speech tag and a frequency. The part-of-speech tag for each nuclear power entity category is shown in Table 1. The frequency of each term in the nuclear power terminology database is manually assigned a value of m.

[0066] Table 1. Parts of Speech Corresponding to Nuclear Power Entity Categories

[0067]

[0068]

[0069] Step 2: Establish a nuclear power plant deprecation term database

[0070] The purpose of building a stop word dictionary is to save storage space and improve the computational efficiency of applications based on word segmentation technology. For example, search engines automatically ignore stop words when indexing pages or processing search requests. The construction of a stop word dictionary includes the following aspects:

[0071] List all punctuation marks as stop words.

[0072] All special characters except "#", uppercase and lowercase letters and numbers are listed as stop words.

[0073] All modal particles are listed as stop words.

[0074] The stop word dictionary consists of fields such as "stop word name" and "creation date".

[0075] Step 3: Establish a thesaurus of nuclear power synonyms

[0076] A thesaurus refers to two words with similar meanings used in the nuclear power industry, such as "leakage," "leak," and "dripping." The creation of a thesaurus involves two aspects: first, manual compilation by nuclear power professionals, followed by importation into the thesaurus; second, automatic thesaurus recognition based on the semantic relationships within the context of given nuclear power corpus using the thesaurus recognition algorithm in step 8, with the resulting words automatically added to the thesaurus. The thesaurus consists of fields such as "Synonym List," "Creation Date," and "Relationships."

[0077] Step 4: Establish a unified thesaurus of pronouns related to nuclear power:

[0078] A lexicon of identical pronouns refers to two words that, when used in the nuclear power industry, refer to the same thing. The process of constructing a lexicon of identical pronouns is as follows:

[0079] This document compiles the equipment codes and names, system numbers and names, building codes and names, room codes and names, Chinese and English names of the nuclear power plant manufacturer, Chinese and English names of the nuclear power plant contractor, material codes and names, and other information from the nuclear power plant's production management system. Other examples include "circulating pump," "circulating water pump," and "CCW pump."

[0080] The equipment code and equipment name of the same device are grouped into the same pronoun and imported into the same pronoun dictionary. For example, the equipment code "9DVN166VD" and the equipment name "electric regulating valve".

[0081] Induce the system number and system name of the same system into the same pronoun, and import them into the same pronoun library. For example, the system number "DVN" and the system name "Nuclear Auxiliary Building Ventilation System".

[0082] Induce the plant code and plant name of the same plant into the same pronoun, and import them into the same pronoun library. For example, the plant code "9WX" and the plant name "Unit 9 - Connection Building".

[0083] Induce the room code and room name of the same room into the same pronoun, and import them into the same pronoun library. For example, the room code "W453" and the room name "Valve and Pipe Room (EBA)".

[0084] Induce the Chinese name and English name of the nuclear power plant manufacturer of the same manufacturer into the same pronoun, and import them into the same pronoun library. For example, the Chinese name of the nuclear power plant manufacturer "Bosch" and the English name of the nuclear power plant manufacturer "BOSCH".

[0085] Induce the Chinese name and English name of the nuclear power plant contractor of the same contractor into the same pronoun, and import them into the same pronoun library. For example, the Chinese name of the nuclear power plant contractor "Leroy Somer" and the English name of the nuclear power plant contractor "LEROY SOMER".

[0086] Induce the material code and material name of the same material into the same pronoun, and import them into the same pronoun library. For example, the material code "9000313362" and the material name "Nitrogen Pressure Gauge [0 - 6 Mpa 2.5 Grade]".

[0087] The same pronoun library consists of fields such as "Synonym List", "Creation Date", "Association Relationship", etc. <​​​Step 6: Machine learning is performed on the nuclear power industry corpus using an encoder-decoder model (where the encoder is a Long Short-Term Memory Neural Network and the decoder is a Conditional Random Field) to automatically identify new words (words not found in the aforementioned "NPP-D" lexicon). Subsequently, a dictionary lookup tree algorithm is used to automatically identify the frequency of new words, and a Hidden Markov Model is used to automatically identify their part of speech. The identified new words are automatically added to the nuclear power professional lexicon. The specific process of new word identification is as follows:

[0090] The encoder model utilizes a Long Short-Term Memory (LSTM) neural network to transform each character of nuclear power-related technical terms into a 768-dimensional feature vector using the BERT framework. These vectors are concatenated to form a feature matrix corresponding to the sentence, which is then fed into the LTM neural network. The LTM network calculates a matrix containing the boundary features of the nuclear power-related technical terms, which is then passed to the decoder. The decoder uses a Conditional Random Field (CRF) to calculate the probability of each character representing a boundary, thus completing word segmentation. Within the segmented sentences, a Hidden Markov Model (HMM) is employed to calculate the part-of-speech tag (POS) of each technical term using statistical conditional probability. After new word recognition, a dictionary lookup tree is constructed for all new words, and these are matched against all documents. The number of matches performed for each word is taken as its frequency.

[0091] Step 7: Supervised machine learning is performed on the nuclear power industry corpus using an encoder-decoder model (where the encoder model is a long short-term memory neural network and the decoder model is a conditional random field). Automatic entity recognition of the nuclear power industry corpus is achieved by annotating and learning the entity categories from a portion of the nuclear power industry corpus. Entity categories are shown in Table 1. New words identified in Step 6 are automatically classified as entities in this step. The automatic entity recognition process is as follows:

[0092] The encoder model utilizes a long short-term memory neural network to transform each character of nuclear power-related technical terms into a 768-dimensional feature vector using the BERT framework. The vectors are concatenated to form a feature matrix corresponding to the sentence. This matrix is ​​then fed into the long short-term memory neural network, which calculates a matrix containing the features of the nuclear power-related technical terms. This feature matrix is ​​then fed into the decoder, where a conditional random field is used to calculate the probability of each character as a boundary, thus completing word segmentation and entity label recognition.

[0093] Step 8: Cluster the "NPP-D" lexicon constructed from the above corpus. Then, for each cluster, use a latent semantic retrieval model and singular value decomposition algorithm to automatically identify synonyms from the nuclear power industry corpus using unsupervised learning. The identified synonyms are then reviewed by professionals and added to the nuclear power synonym lexicon. The automatic synonym identification process is as follows:

[0094] For each word in the NPP-D vocabulary, a semantic vector is calculated using a word embedding model. This vector is then clustered using the k-Means clustering algorithm to obtain several clusters. Within each cluster, the word feature vectors are concatenated to construct a feature matrix. This feature matrix is ​​then decomposed into three distinct matrices: a left matrix, a singular value diagonal matrix, and a right matrix, using singular value decomposition (SVD). The values ​​in the singular value diagonal matrix are analyzed, and singular values ​​exceeding a certain threshold are filtered out. The corresponding columns in the left matrix are then searched, and the word pairs with the closest similarity are selected as synonyms.

[0095] Step 9: Cluster the “NPP-D” lexicon constructed from the above corpus. Then, on each cluster, use a latent semantic retrieval model and singular value decomposition algorithm to automatically identify identical pronouns in the nuclear power industry corpus using unsupervised learning. The identified identical pronouns are then reviewed by professionals and added to the nuclear power identical pronoun lexicon. The automatic identification process for identical pronouns is as follows:

[0096] For each word in the NPP-D vocabulary, a semantic vector is calculated using a word embedding model. These vectors are then clustered using the k-Means clustering algorithm to obtain several clusters. Within each cluster, the word feature vectors are concatenated to construct a feature matrix. This feature matrix is ​​then decomposed into three distinct matrices: a left matrix, a singular value diagonal matrix, and a right matrix, using singular value decomposition. The right matrix, after decomposition, represents the semantic similarity matrix. Word pairs with high semantic similarity are designated as sharing the same pronoun.

[0097] Step 10: Based on the “NPP-D” lexicon, a nuclear power word segmentation model is constructed using a dynamic programming algorithm to find the maximum probability path and a hidden Markov model in an unsupervised learning manner.

[0098] The Hidden Markov Model assigns the frequency of each word in the existing nuclear power corpus as its probability of occurrence, and sets the conditional probability of new words to the minimum frequency. The existing nuclear power terminology is trained unsupervised using the model to obtain a probability matrix for each sentence. The matrix length corresponds to the sentence length, and the width corresponds to the label category. A dynamic programming pathfinding algorithm is used to calculate the path with the highest overall probability from the probability matrix. Based on the matrix row and column indices traversed by the path, BMES (Begin-Middle-End-Single) labeling is applied to mark the boundaries of each word, and the corresponding label is used as the character in the sentence.

[0099] Step 11: Based on the nuclear power plant word segmentation model, develop a front-end interface for nuclear power plant word segmentation and establish a nuclear power plant word segmentation tool. The nuclear power plant word segmentation tool achieves accurate Chinese word segmentation of nuclear power plant corpora by calling the nuclear power plant word segmentation model in Step 10. The nuclear power plant professional word segmentation tool is described in English as Nuclear Power Plant Word Segmentation Toolkit, abbreviated as "NPP-WST".

Claims

1. A professional word segmentation method applied to the nuclear power industry, characterized in that, Includes the following steps: Step 1: Establish a nuclear power-related thesaurus by analyzing the corpus; Step 2: Establish a nuclear power plant delimited word database; Step 3: Establish a thesaurus of nuclear power synonyms; Step 4: Establish a unified thesaurus of pronouns for nuclear power; Step 5: Combine the nuclear power professional terminology database, nuclear power stop word database, nuclear power synonym database, nuclear power common pronoun database, and jieba general terminology database constructed in steps 1 to 4 above to form a nuclear power terminology database; Step 6: By employing an encoder-decoder model to perform machine learning on the nuclear power industry corpus, automatic new word recognition can be achieved on the nuclear power industry corpus; In step 6, a dictionary lookup tree algorithm is used to automatically identify the frequency of new words, and a hidden Markov model is used to automatically identify the part of speech of new words. The identified new words are automatically included in the nuclear power professional thesaurus. The specific process of new word identification is as follows: The encoder model utilizes a long short-term memory neural network to transform each character of nuclear power-related technical terms into a 768-dimensional feature vector using the BERT framework. These vectors are then concatenated to form a feature matrix corresponding to the sentence. This matrix is ​​fed into the long short-term memory neural network, which calculates a matrix containing the boundary features of the nuclear power-related technical terms. This boundary feature matrix is ​​then fed into the decoder, where a conditional random field is used to calculate the probability of each character representing a boundary, thus completing word segmentation. Simultaneously, a hidden Markov model is employed in the segmented sentences to calculate the part-of-speech tag for each technical term using statistical conditional probability. After new word recognition, a dictionary lookup tree is constructed for all new words, and the results are matched against all documents. The number of matches completed for each word is taken as the term frequency of the new word. Step 7: Supervised machine learning is performed on the nuclear power industry corpus using an encoder-decoder model. By annotating and learning the nuclear power entity categories from a portion of the nuclear power industry corpus, automatic entity recognition of the nuclear power industry corpus is achieved. Step 8: Cluster the “NPP-D” lexicon constructed from the above corpus, and then use the latent semantic retrieval model and singular value decomposition algorithm on each cluster to achieve automatic synonym recognition of nuclear power industry corpus in an unsupervised learning manner. The identified synonyms are included in the nuclear power synonym lexicon after being reviewed by professionals. Step 9: Cluster the “NPP-D” lexicon constructed from the above corpus, and then use the latent semantic retrieval model and singular value decomposition algorithm on each cluster to automatically identify the same pronouns in the nuclear power industry corpus in an unsupervised learning manner. The identified same pronouns are included in the nuclear power same pronoun lexicon after being reviewed by professionals. Step 10: Based on the "NPP-D" lexicon, a nuclear power word segmentation model is constructed using a dynamic programming algorithm to find the maximum probability path and a hidden Markov model in an unsupervised learning manner; Step 11: Based on the nuclear power word segmentation model, establish a nuclear power word segmentation tool. The nuclear power word segmentation tool achieves accurate Chinese word segmentation of nuclear power corpus by calling the nuclear power word segmentation model in Step 10.

2. The professional word segmentation method applied to the nuclear power industry according to claim 1, wherein: Step 1 involves compiling various types of nuclear power terminology by searching through system design manuals, equipment delivery manuals, structure layout design documents, equipment failure mode documents, as well as equipment lists, material lists, organizational lists, personnel information lists, and risk lists in the nuclear power plant's production management system.

3. The professional word segmentation method applied to the nuclear power industry according to claim 1, characterized in that: Step 2, the construction of the stop words, includes the following aspects: List all punctuation marks as stop words; All special characters except "#", uppercase and lowercase letters and numbers are listed as stop words; List all modal particles as stop words; The stop word dictionary consists of the fields "stop word name" and "creation date".

4. The professional word segmentation method applied to the nuclear power industry according to claim 1, wherein: In step 4, the term "same pronoun lexicon" refers to two words that, when used in the nuclear power industry, refer to the same thing. The process of constructing the same pronoun lexicon is as follows: The device code and device name of the same device are grouped into the same pronoun and imported into the same pronoun dictionary; The system number and system name of the same system are grouped into the same pronoun and imported into the same pronoun dictionary; The factory codes and factory names of the same factory building are grouped into the same pronoun and imported into the same pronoun dictionary; Group the room codes and room names of the same room into the same pronoun and import them into the same pronoun dictionary; The Chinese and English names of nuclear power plant manufacturers from the same manufacturer are grouped into the same pronoun and imported into the same pronoun dictionary. The Chinese and English names of nuclear power plant contractors under the same contractor are grouped into the same pronoun and imported into the same pronoun dictionary. The material codes and material names of the same material are grouped into the same pronoun and imported into the same pronoun thesaurus; The same pronoun dictionary consists of the fields of "synonym list", "creation date" and "association".

5. A professional word segmentation method for the nuclear power industry as described in claim 1, characterized in that: The automatic entity recognition process in step 7 is as follows: The encoder model utilizes a long short-term memory neural network to transform each character of nuclear power-related technical terms into a 768-dimensional feature vector using the BERT framework. The vectors are concatenated to form a feature matrix corresponding to the sentence. This matrix is ​​then fed into the long short-term memory neural network, which calculates a matrix containing the features of the nuclear power-related technical terms. This feature matrix is ​​then fed into the decoder, where a conditional random field is used to calculate the probability of each character as a boundary, thus completing word segmentation and entity label recognition.

6. The professional word segmentation method applied to the nuclear power industry according to claim 1, wherein: The automatic synonym recognition process in step 8 is as follows: For each word in the NPP-D vocabulary, a semantic vector is calculated using a word embedding model. The vectors are then clustered into several clusters using the k-Means clustering algorithm. Within each cluster, the word feature vectors are concatenated to construct a feature matrix. The feature matrix is ​​then decomposed into three different matrices: a left matrix, a singular value diagonal matrix, and a right matrix, using a singular value decomposition algorithm. The values ​​in the singular value diagonal matrix are analyzed, and singular values ​​exceeding a certain threshold are filtered out. The corresponding columns in the left matrix are then searched, and the word pairs with the closest similarity are selected as synonyms.

7. The professional word segmentation method applied to the nuclear power industry according to claim 1, wherein: The automatic identification process for the same pronoun in step 8 is as follows: For each word in the NPP-D vocabulary, a semantic vector is calculated using a word embedding model. The vectors are then clustered into several clusters using the k-Means clustering algorithm. Within each cluster, the word feature vectors are concatenated to construct a feature matrix. The feature matrix is ​​then decomposed into three different matrices: a left matrix, a singular value diagonal matrix, and a right matrix using a singular value decomposition algorithm. The right matrix after decomposition is the semantic similarity matrix. Word pairs with high semantic similarity are set as the same pronoun.

8. The professional word segmentation method applied to the nuclear power industry according to claim 1, wherein: In step 10, the hidden Markov model marks the frequency of each word in the existing nuclear power corpus as its probability of occurrence, sets the conditional probability of new words to the minimum frequency, and trains the existing nuclear power professional vocabulary in an unsupervised manner to obtain the probability matrix of the sentences. The length of the matrix corresponds to the length of the sentences, and the width corresponds to the label category. Through a dynamic programming pathfinding algorithm, a path with the highest comprehensive probability is calculated in the probability matrix. Based on the matrix row and column numbers traversed by the path, Begin-Middle-End-Single marking is performed to mark the boundary of each word, and the corresponding label is used as each character in the sentence.

Images