Term dynamic updating method for electric power professional questions and answers

Abstract

The invention relates to the technical field of computer information processing, discloses a power professional question and answer term dynamic updating method, and aims to solve the problem of semantic inconsistency of a question and answer system caused by updating lag of a traditional static term library. The method comprises the following steps: constructing a power professional term knowledge base and establishing a term semantic association network; dynamically monitoring new terms through a web crawler and a dual-trigger mechanism; semantic verification and disambiguation are performed in combination with the language model and the knowledge graph; updating the term library and maintaining a change history by adopting a version-controlled incremental algorithm; and synchronously triggering incremental learning and model fine tuning of a semantic understanding module of the question-answering system to realize collaborative updating of terms and the question-answering system. According to the technical scheme, the term change of the power industry can be perceived in real time, the semantic consistency of the question-answering system is ensured, the adaptive capacity and service efficiency of the system are remarkably improved, and the manual maintenance cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of computer information processing technology, specifically to a method for dynamically updating terminology in electrical engineering Q&A. Background Technology

[0002] With the deepening application of artificial intelligence (AI) technology in the power industry, its value in the design review of power grid infrastructure projects is becoming increasingly prominent, serving as a key means to improve review efficiency, ensure review quality, optimize resource allocation, and support scientific investment decisions. Currently, power system design review has gradually transformed from a traditional offline model to an information-based and digital platform. Relying on intelligent rule bases and structured data processing capabilities, it has initially achieved consistency verification and automated analysis of feasibility study reports, construction drawings, material selection, and cost data. However, the power industry terminology system is vast and dynamically evolving, covering multiple dimensions such as equipment naming, technical specifications, standard clauses, and regional expressions. Its semantic consistency and timeliness directly affect the accuracy of AI models' understanding of design documents and the reliability of review logic.

[0003] Among them, the power industry Q&A system, as a crucial interactive interface supporting intelligent review, needs to accurately understand the engineering questions raised by users and return technical answers that conform to the latest specifications. This system heavily relies on the completeness and timeliness of the underlying terminology knowledge base, and its performance is directly limited by the breadth and speed of terminology coverage. Existing Q&A systems mostly use static terminology dictionaries or periodically manually maintained knowledge bases, which are difficult to adapt to the real-world needs of frequent revisions to power grid standards, updates to local regulations, and the rapid emergence of new equipment terminology.

[0004] In existing technologies, terminology updates typically rely on manual compilation and batch import, which suffers from problems such as slow response times, high maintenance costs, and weak semantic connections. This leads to recognition failures, matching errors, and even logical misjudgments when faced with new or variant terms. Furthermore, the lack of dynamic awareness of terminology context makes it difficult for the system to distinguish between homonyms or handle abbreviations, colloquial expressions, and other non-standard expressions. In addition, in multi-round question-and-answer sessions or complex engineering scenarios, static terminology databases cannot achieve adaptive evolution and contextual alignment of terminology meanings, severely impacting question-and-answer accuracy and review assistance effectiveness. Therefore, there is an urgent need for a question-and-answer support method that can achieve dynamic awareness, automatic updating, and semantic collaboration of power industry terminology to ensure the continuous and accurate service capability of intelligent review systems in rapidly changing technical specification environments. Summary of the Invention

[0005] The purpose of this invention is to provide a method for dynamically updating terminology in electrical engineering Q&A, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for dynamically updating terminology in electrical engineering Q&A includes the following specific steps: Step 1: Construct a knowledge base of power industry terminology. Obtain power industry standard documents, technical specifications, equipment manuals and engineering case materials through multi-source heterogeneous data acquisition interfaces. Use a named entity recognition model based on a combination of bidirectional long short-term memory network and conditional random field to extract power industry terminology entities and establish a semantic association network between terms. Step 2: Establish a terminology dynamic monitoring mechanism, configure a web crawler system to continuously monitor the websites of the power industry's competent authorities, standard update announcements, and technical forum dynamics, and set dual trigger conditions based on keyword matching and semantic similarity calculation. When potential new terms or term change information is detected, the update process will be automatically started. Step 3: Implement semantic verification and disambiguation of terms. Analyze the context of the candidate terms obtained from monitoring. Calculate the semantic similarity between the candidate terms and related terms in the existing terminology database using a pre-trained language model. Combine this with a knowledge graph of the power industry to perform semantic conflict detection and disambiguation. Step 4: Perform incremental updates to the terminology knowledge base. The verified terms and their semantic relationships are structured and processed. A version control-based incremental update algorithm is used to integrate new terms into the existing knowledge base, while maintaining the terminology change history and version tracking information. Step 5: Implement collaborative updating of terminology in the question-answering system. After the terminology knowledge base is updated, the model retraining process of the semantic understanding module of the question-answering system is automatically triggered. The terminology embedding vector and semantic matching model are updated using an incremental learning method to ensure that the question-answering system and the new terminology database maintain semantic consistency.

[0007] Preferably, the multi-source heterogeneous data acquisition interface in step 1 supports at least four data formats, including text files, database tables, web page content, and application programming interfaces. The data acquisition frequency can be dynamically adjusted according to the data source update cycle, with a minimum acquisition interval of 1 minute.

[0008] Preferably, the named entity recognition model adopts a 12-layer bidirectional long short-term memory network structure with a hidden layer dimension of 768. The conditional random field label set includes 15 types of power professional entities such as equipment name, technical parameters, and standard number. The entity recognition accuracy on the power field test set reaches 97.5%.

[0009] Preferably, in step 2, the web crawler system is configured with a distributed architecture, containing at least 3 crawler nodes, each node is equipped with an independent Internet Protocol address pool, supports concurrent processing of up to 128 data sources, and can process more than 1 million web page data per day.

[0010] Preferably, the semantic similarity calculation employs a deep semantic matching model continuously trained on a corpus from the power sector, with a similarity threshold set at 0.85. When the semantic similarity between a candidate term and existing terms is lower than this threshold, the system automatically marks it as a potential new term. The semantic similarity calculation formula is as follows:

[0011] in, and Terms and Embedding vectors in pre-trained language models in the power sector This represents the cosine similarity between the two.

[0012] Preferably, the contextual analysis in step 3 employs an attention-enhanced context encoder, which can capture the semantic changes of terms in specific engineering scenarios and achieves an accuracy rate of over 99% in distinguishing homonymous terms.

[0013] Preferably, the pre-trained language model is based on the Transformer architecture, contains 24 encoder layers, and has 340 million parameters. It outperforms general language models by 30% on power industry text understanding tasks.

[0014] Preferably, the semantic conflict detection and disambiguation process includes: establishing a semantic conflict detection rule base for terms, which includes semantic constraints based on the ontology of the power field; when a semantic conflict is detected, activating a multi-evidence fusion disambiguation algorithm, which comprehensively considers the frequency of term occurrence, contextual consistency, and domain expert annotation information.

[0015] Preferably, the incremental update algorithm based on version control in step 4 employs an improved binary search and hash indexing technique to ensure that the average response time of a single term update operation is less than 50 milliseconds in a knowledge base containing more than 10 million term records.

[0016] Preferably, the terminology change history is stored in a time-series database to retain a complete data change trajectory and support terminology status retrospection and change impact analysis at any point in time.

[0017] Preferably, the model retraining process in step 5 employs a dynamic learning rate adjustment strategy. The initial learning rate is set to 0.001, and it decays exponentially with each training epoch, ensuring that the model retains its existing knowledge while absorbing new terminology. The elastic weight consolidation (EWC) algorithm is used during incremental learning, and its loss function is defined as follows:

[0018] in, For the task loss on the new terminology data, These are the current model parameters. For the old model parameters, These are the diagonal elements of the parametric importance matrix. is the regularization coefficient.

[0019] Preferably, the update process of the semantic understanding module includes online learning of term embedding vectors and fine-tuning of the parameters of the semantic matching model. The entire update process can be completed within 2 hours, minimizing the impact on the normal service of the question-answering system.

[0020] Preferably, it also includes a terminology update quality assessment mechanism, which automatically extracts 5% of the update records for manual review after each terminology database update, and optimizes the parameter settings of the terminology verification and disambiguation algorithm based on the review results.

[0021] Preferably, the method supports term update rollback functionality, which allows for rapid restoration to the previous stable version when an update error or semantic inconsistency is detected, with a maximum rollback time of no more than 5 minutes.

[0022] Preferably, the terminology dynamic update method is integrated into the power intelligent review platform, which provides terminology services to multiple business systems through a standardized application programming interface, and processes more than 100,000 terminology query requests per day.

[0023] Compared with the prior art, the beneficial effects achieved by the present invention are: By constructing a power industry terminology knowledge base with dynamic monitoring and automatic updates, real-time perception and rapid response to changes in power industry terminology are achieved, effectively solving the problem of lagging updates caused by traditional static terminology databases. Deep learning-based terminology extraction and semantic verification technologies ensure the accuracy and semantic consistency of new terminology identification, while incremental and collaborative update mechanisms guarantee real-time synchronization between the question-and-answer system and the terminology database. This method significantly improves the adaptability and service level of the power industry question-and-answer system in a rapidly changing technological environment, provides reliable terminology support for intelligent review systems, and substantially reduces manual maintenance costs and response delays. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall technical solution architecture of the method of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the term dynamic monitoring and triggering mechanism in this invention; Figure 3 This is a flowchart illustrating the logical flow of terminological semantic verification and disambiguation in this invention. Figure 4 This is a logical flowchart of the incremental update and version control of the terminology knowledge base in this invention; Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow between the collaborative updating of terms and incremental learning of the model in the question-answering system of this invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0027] Currently, with the deepening application of artificial intelligence in the power industry, the design review of power grid infrastructure projects is accelerating its transformation towards information and digital platforms. Intelligent rule bases and structured data processing capabilities have initially supported the consistency verification and automated analysis of feasibility study reports, construction drawings, material selection, and cost data. However, the power industry terminology system is vast and dynamically evolving, encompassing multiple dimensions such as equipment naming, technical specifications, standard clauses, and regional expressions. Its semantic consistency and timeliness directly affect the accuracy of the AI ​​model's understanding of design documents and the reliability of the review logic. As a key interactive interface for intelligent review, the power industry question-and-answer system heavily relies on the completeness and timeliness of the underlying terminology knowledge base. Existing systems often employ static terminology dictionaries or periodically manually maintained knowledge bases, which struggle to adapt to the frequent revisions of power grid standards, updates to local regulations, and the rapid emergence of new equipment terminology. This leads to terminology recognition failures, matching deviations, and even logical misjudgments. Furthermore, they lack the ability to dynamically perceive the context of terms, failing to effectively handle homonymous terms, abbreviations, or colloquial expressions. In multi-round question-and-answer sessions or complex engineering scenarios, they also cannot achieve adaptive evolution and context alignment of terminology meanings. To address the aforementioned technical issues, this invention proposes a method for dynamically updating terminology in power industry Q&A. By constructing a terminology knowledge base system with dynamic monitoring, automatic verification, incremental updates, and collaborative synchronization capabilities, it achieves real-time perception and accurate response to changes in power industry terminology, and applies this method to the dynamic updating of terminology in power industry Q&A.

[0028] refer to Figure 1The overall technical architecture of the dynamic terminology update method for power industry Q&A proposed in this invention includes a terminology knowledge base construction module, a terminology dynamic monitoring module, a terminology semantic verification and disambiguation module, a terminology knowledge base incremental update module, and a Q&A system terminology collaborative update module. Each module achieves tight coupling and efficient collaboration through standardized data interfaces and event-driven mechanisms. In the above-mentioned dynamic terminology update method for power industry Q&A, step (1) involves constructing a power industry terminology knowledge base by acquiring power industry standard documents, technical specification documents, equipment manuals, and engineering case data through a multi-source heterogeneous data acquisition interface. A named entity recognition model based on a combination of bidirectional long short-term memory networks and conditional random fields is used to extract power industry terminology entities and establish a semantic association network between terms. Specifically, the multi-source heterogeneous data acquisition interface supports at least four data formats, including text files, database tables, web page content, and application programming interfaces. The data acquisition frequency can be dynamically adjusted according to the data source update cycle, with a minimum acquisition interval of 1 minute, ensuring minute-level response capability for high-timeliness data sources (such as standard announcement websites). The collected raw data is uniformly encoded, converted, and format-standardized before being input into the named entity recognition model. The model employs a 12-layer bidirectional long short-term memory network structure with 768 hidden layer dimensions. The conditional random field label set includes 15 power industry entity types, such as equipment names, technical parameters, and standard numbers. On a power industry test set, the entity recognition accuracy reaches 97.5%. The model output is a structured list of terminology entities, with each entity accompanied by its type label, text span, and confidence score. Subsequently, based on terminology co-occurrence frequency, syntactic dependency relations, and domain ontology constraints, a semantic association network between terms is constructed. This network is stored in a graph structure, where nodes represent terminology entities, and edges represent semantic relationships (such as synonymy, hyponymy, and attribute association). Edge weights are determined by a relation strength quantification index, providing a foundation for subsequent semantic similarity calculation and conflict detection.

[0029] In the above-mentioned method for dynamically updating terminology in power industry Q&A, step (2) establishes a dynamic terminology monitoring mechanism. This involves configuring a web crawler system to continuously monitor the websites of power industry regulatory authorities, standard update announcements, and technical forum updates. A dual triggering condition based on keyword matching and semantic similarity calculation is set up, automatically initiating the update process when potential new terms or terminology change information is detected. Specifically, the web crawler system is configured with a distributed architecture, containing at least three crawler nodes. Each node is equipped with an independent Internet Protocol address pool, supporting concurrent processing of up to 128 data sources. It can process over 1 million web page data daily, effectively circumventing anti-crawler strategies and ensuring the stability and coverage of data collection. The crawler task scheduler dynamically allocates collection tasks according to a preset monitoring target list (including the National Energy Administration, China Electricity Council, and official websites of various provincial power grid companies, etc.), and performs incremental crawling and deduplication of web page content. After text cleaning, the crawled content is first initially screened using a keyword matching engine. The keyword library covers high-frequency trigger words such as "new standard," "revision," "abolition," "new equipment," and "technical specification update." Secondly, for the selected candidate text fragments, a deep semantic matching model is used to calculate their semantic similarity with terms in the existing terminology library. This semantic similarity calculation employs a deep semantic matching model continuously trained on a corpus from the power industry, with a similarity threshold set at 0.85. When the semantic similarity between a candidate term and existing terms is lower than this threshold, the system automatically marks it as a potential new term. This semantic similarity calculation process is precisely quantified by the following formula:

[0030] in, and Terms and Embedding vectors in pre-trained language models in the power sector This represents the cosine similarity between the two terms. The dual-trigger design ensures high sensitivity and low false alarm rate in the monitoring mechanism. A term update event is generated only when a keyword match is found and the semantic similarity is below a threshold. The candidate term and its context fragment are then stored in a queue for verification. (See [link to relevant documentation]). Figure 2 The core principle framework of the term dynamic monitoring and triggering mechanism shown is as follows.

[0031] In the above-mentioned dynamic terminology update method for power industry question answering, step (3) involves performing semantic verification and disambiguation processing of terms. This involves analyzing the context of candidate terms obtained through monitoring, calculating the semantic similarity between candidate terms and related terms in the existing terminology database using a pre-trained language model, and combining this with a knowledge graph of the power industry for semantic conflict detection and disambiguation. Specifically, the context analysis employs an attention-enhanced context encoder, which can capture semantic changes of terms in specific engineering scenarios and achieves an accuracy rate exceeding 99% in distinguishing homonymous terms. This encoder extracts context windows of 50 words before and after each candidate term, dynamically weighting the importance of different context words through a multi-head attention mechanism to generate context vectors rich in scene semantics. The pre-trained language model is based on the Transformer architecture, containing 24 encoder layers and 340 million parameters. Its performance on power industry text understanding tasks is 30% better than general language models, and its output embedding vectors accurately reflect the professional semantics of terms in the power industry. Semantic conflict detection and disambiguation processing includes: establishing a rule base for semantic conflict detection of terms, containing semantic constraints based on the ontology of the power industry (such as "transformer" and "circuit breaker" cannot be synonymous, and "rated voltage" must be associated with a specific equipment type); when a semantic conflict is detected, a multi-evidence fusion disambiguation algorithm is activated, comprehensively considering the frequency of term occurrence, contextual consistency, and domain expert annotation information. For example, for the candidate term "GIS," the system analyzes whether its context involves keywords such as "gas-insulated" and "substation." If the context points to "Geographic Information System," it is determined to be a non-power industry term and filtered out; if it points to "gas-insulated switchgear," it further verifies its semantic consistency with the existing term "gas-insulated metal-enclosed switchgear" and decides whether to establish a synonym relationship or introduce it as a new term. The entire verification process is carried out within... Figure 3 The process is executed within the logical framework of semantic verification and disambiguation of the terms shown, ensuring the accuracy and semantic purity of the new terms.

[0032] In the above-mentioned dynamic terminology update method for power industry Q&A, step (4) involves performing incremental updates to the terminology knowledge base, structuring the verified terms and their semantic relationships, and integrating new terms into the existing knowledge base using a version-controlled incremental update algorithm, while maintaining the terminology change history and version tracking information. Specifically, the version-controlled incremental update algorithm uses improved binary search and hash index technology to ensure that the average response time of a single term update operation is less than 50 milliseconds in a knowledge base containing more than 10 million terminology records. New terms and their relationships are encapsulated into a structured update package, which includes terminology ID, text, type, semantic vector, list of associated terms, and metadata. The update algorithm first locates the insertion or update position in the terminology master index, uses hash index to accelerate the search, and then performs atomic write operations to ensure data consistency. At the same time, the terminology change history is stored in a time-series database to retain a complete data change trajectory. Each record includes a change timestamp, operation type (add, modify, delete), operator (system or manual), state before change, and state after change, supporting terminology state backtracking and change impact analysis at any point in time. For example, it's possible to trace the entire process of a term from its initial introduction to subsequent semantic revisions and assess its impact on historical question-and-answer records. This process is as follows: Figure 4 As shown, the logical steps of incremental updates and version control of the terminology knowledge base are clearly demonstrated.

[0033] In the above-mentioned dynamic terminology update method for power industry Q&A, step (5) realizes the collaborative update of terminology in the Q&A system. After the terminology knowledge base is updated, the model retraining process of the semantic understanding module of the Q&A system is automatically triggered. The terminology embedding vector and semantic matching model are updated using an incremental learning method to ensure that the Q&A system and the new terminology database maintain semantic consistency. Specifically, the model retraining process adopts a dynamic learning rate adjustment strategy. The initial learning rate is set to 0.001 and decreases exponentially with the increase of training rounds to ensure that the model retains the memory of the original knowledge while absorbing new terminology knowledge. The elastic weight consolidation (EWC) algorithm is used in the incremental learning process, and its loss function is defined as follows:

[0034] in, For the task loss on the new terminology data, These are the current model parameters. For the old model parameters, These are the diagonal elements of the parametric importance matrix. is the regularization coefficient.

[0035] This algorithm effectively mitigates the catastrophic forgetting problem by penalizing significant modifications to important parameters of old tasks. The update process of the semantic understanding module includes online learning of term embedding vectors and fine-tuning of the semantic matching model parameters. The entire update process can be completed within 2 hours, minimizing the impact on the normal service of the question-answering system. After the update is complete, the system automatically performs a service switch, deploying the new model to the production environment and ensuring service stability through health checks. This collaborative update mechanism is as follows: Figure 5 As shown, this illustrates the multi-level interaction and data flow between collaborative terminology updates and incremental model learning in a question-answering system.

[0036] To further enhance system robustness and service quality, this invention also includes a terminology update quality assessment mechanism. After each terminology database update, 5% of the update records are automatically sampled for manual review. Based on the review results, the parameter settings of the terminology verification and disambiguation algorithms are optimized, forming a closed-loop feedback. Furthermore, the method supports terminology update rollback functionality. When update errors or semantic inconsistencies are detected (such as a sudden drop in question-and-answer accuracy or errors discovered through manual review), the system can quickly revert to the previous stable version, with a maximum rollback time of no more than 5 minutes, ensuring the continuity and reliability of system services. This dynamic terminology update method is integrated into the power intelligent review platform, providing terminology services to multiple business systems through standardized application programming interfaces. It processes over 100,000 terminology query requests daily, effectively supporting multiple core business scenarios such as intelligent review, design verification, and material selection.

[0037] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for dynamically updating terminology in a question-and-answer format for the power industry, characterized in that: The specific steps include the following: Step 1: Construct a knowledge base of power industry terminology. Obtain power industry standard documents, technical specifications, equipment manuals and engineering case materials through multi-source heterogeneous data acquisition interfaces. Use a named entity recognition model based on a combination of bidirectional long short-term memory network and conditional random field to extract power industry terminology entities and establish a semantic association network between terms. Step 2: Establish a terminology dynamic monitoring mechanism, configure a web crawler system to continuously monitor the websites of the power industry's competent authorities, standard update announcements, and technical forum dynamics, and set dual trigger conditions based on keyword matching and semantic similarity calculation. When potential new terms or term change information is detected, the update process will be automatically started. Step 3: Implement semantic verification and disambiguation of terms. Analyze the context of the candidate terms obtained from monitoring. Calculate the semantic similarity between the candidate terms and related terms in the existing terminology database using a pre-trained language model. Combine this with a knowledge graph of the power industry to perform semantic conflict detection and disambiguation. Step 4: Perform incremental updates to the terminology knowledge base. The verified terms and their semantic relationships are structured and processed. A version control-based incremental update algorithm is used to integrate new terms into the existing knowledge base, while maintaining the terminology change history and version tracking information. Step 5: Implement collaborative updating of terminology in the question-answering system. After the terminology knowledge base is updated, the model retraining process of the semantic understanding module of the question-answering system is automatically triggered. The terminology embedding vector and semantic matching model are updated using an incremental learning method to ensure that the question-answering system and the new terminology database maintain semantic consistency.

2. The method for dynamically updating terminology in a power industry Q&A system according to claim 1, characterized in that: The multi-source heterogeneous data acquisition interface supports at least four data formats, including text files, database tables, web page content, and application programming interfaces. The data acquisition frequency can be dynamically adjusted according to the data source update cycle, with a minimum acquisition interval of 1 minute.

3. The method for dynamically updating terminology in a power industry Q&A system according to claim 1, characterized in that: The named entity recognition model adopts a 12-layer bidirectional long short-term memory network structure with a hidden layer dimension of 768. The conditional random field label set includes 15 types of power professional entities such as equipment name, technical parameters, and standard number.

4. The method for dynamically updating terminology in a power industry Q&A system according to claim 1, characterized in that: The web crawler system is configured with a distributed architecture, containing at least 3 crawler nodes, each node is equipped with an independent Internet Protocol address pool, and supports concurrent processing of up to 128 data sources.

5. The method for dynamically updating terminology in a power industry Q&A system according to claim 1, characterized in that: The semantic similarity calculation adopts a deep semantic matching model that is continuously trained based on corpus in the power field. The similarity threshold is set to 0.

85. When the semantic similarity between a candidate term and an existing term is lower than this threshold, the system automatically marks it as a potential new term.

6. The method for dynamically updating terminology in a power industry Q&A system according to claim 1, characterized in that: The contextual analysis employs an attention-enhanced context encoder, which can capture the semantic changes of terms in specific engineering scenarios and achieves an accuracy of over 99% in distinguishing homonymous terms.

7. The method for dynamically updating terminology in a power industry Q&A system according to claim 1, characterized in that: The pre-trained language model is based on the Transformer architecture, contains 24 encoder layers, and has 340 million parameters.

8. The method for dynamically updating terminology in a power industry Q&A system according to claim 1, characterized in that: The semantic conflict detection and disambiguation process includes: establishing a semantic conflict detection rule base for terms, which includes semantic constraints based on the ontology of the power domain; when a semantic conflict is detected, a multi-evidence fusion disambiguation algorithm is activated, which comprehensively considers the frequency of term occurrence, contextual consistency, and domain expert annotation information.

9. The method for dynamically updating terminology in a power industry Q&A system according to claim 1, characterized in that: The version-controlled incremental update algorithm employs improved binary search and hash indexing techniques to ensure that the average response time for a single term update operation is less than 50 milliseconds in a knowledge base containing more than 10 million term records. The term change history is stored in a time-series database, supporting term status backtracking and change impact analysis at any point in time.

10. The method for dynamically updating terminology in a power industry Q&A system according to claim 1, characterized in that: The model retraining process adopts a dynamic learning rate adjustment strategy, with the initial learning rate set to 0.001 and decreasing exponentially with the number of training rounds. The incremental learning method is based on the elastic weight consolidation algorithm, which calculates the importance matrix of the model parameters and applies regularization constraints to the important parameters during training. The update process of the semantic understanding module includes online learning of term embedding vectors and fine-tuning of the parameters of the semantic matching model. The entire update process is completed within 2 hours.

Images