Key element extraction method oriented to automatic generation of cognitive electronic warfare test scheme

Abstract

The invention discloses a key element extraction method oriented to automatic generation of a cognitive electronic warfare test scheme. The key element extraction method comprises the following steps: acquiring a cognitive electronic warfare demand document data set; obtaining a candidate element set based on the demand document data set in combination with task perception template matching and position constraint priority rules; performing semantic sorting and screening on the candidate element set according to hierarchical semantic representation to obtain screened candidate elements; processing the screened candidate elements according to a domain term knowledge base to obtain normalized elements; and obtaining structured key element information based on the normalized elements. According to the method, the unstructured demand text can be converted into the structured quintuple capable of being used for automatic generation of the test scheme in a high-quality mode, coverage, semantic accuracy and term normalization are considered, and engineering deployment and manual review are facilitated.

Description

Technical Field

[0001] This invention belongs to the field of cognitive electronic warfare technology, and in particular relates to a method for extracting key elements for the automatic generation of cognitive electronic warfare test schemes. Background Technology

[0002] With the rapid evolution of information warfare, cognitive electronic warfare equipment is exhibiting unprecedented complexity and diversity in intelligent decision-making, adaptive countermeasures, and multi-domain collaboration. Traditional test scheme design mainly relies on expert manual analysis, gradually sorting out and repeatedly revising operational missions, test elements, and test activities. However, facing scenarios that require comprehensive evaluation of equipment performance in realistic, complex, and diverse environments, the number of test elements is growing exponentially. This makes it difficult for experts to quickly, comprehensively, and accurately identify all key test elements from requirements documents, resulting in longer scheme development cycles, increased costs, and increased risks of missed tests and inconsistencies.

[0003] Currently, there are two main technical approaches to intelligent design of experimental plans: one is based on knowledge graphs, which constructs a structured domain knowledge system, uses key elements in the requirements as anchors to locate and activate relevant knowledge subgraphs, and then uses generative models to synthesize preliminary plans; the other is based on large-scale language models, which leverage their powerful natural language processing capabilities to output experimental plans through techniques such as prompting engineering and retrieval enhancement. While both approaches have shown some feasibility, they still have significant limitations: knowledge graph methods are highly dependent on the completeness and accuracy of domain knowledge; and large language models are prone to generating content "illusions" when applied in specialized domains, producing information that does not conform to reality. In-depth analysis shows that accurately invoking background knowledge highly depends on the precise identification of key elements in the requirements document; simultaneously, consistency verification by matching domain knowledge with key elements can effectively alleviate the illusion problem of large language models. Therefore, achieving efficient and accurate extraction of key elements in the requirements has become the core link in improving the quality and reliability of generated experimental plans.

[0004] Existing methods also have shortcomings in the field of key element extraction. Graph-based ranking methods (such as TextRank and SingleRank) can utilize word co-occurrence structure scoring, but graphs constructed based on shallow statistics struggle to capture cross-sentence semantic dependencies, resulting in limited effectiveness for long documents. Methods based on pre-trained language models (such as SIFRank and MDERank) offer improvements in semantic understanding, but typically employ a single semantic modeling strategy, lacking effective integration of domain knowledge and engineering standards, leading to deficiencies in terminology standardization and usability of the extraction results. Furthermore, most existing methods tend towards end-to-end single-stage designs, attempting to simultaneously achieve coverage, accuracy, and standardization within a unified framework. However, these objectives often constrain each other, making it difficult to achieve ideal results simultaneously in complex engineering scenarios. In conclusion, current technology urgently needs a new method that integrates domain semantics and knowledge constraints while ensuring high coverage, achieving high accuracy and high standardization in key element extraction. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention proposes a key element extraction method for the automatic generation of cognitive electronic warfare test schemes. This invention can convert unstructured requirement text into structured quintuples that can be used for the automatic generation of test schemes with high quality, taking into account coverage, semantic accuracy, and terminology standardization, and facilitating engineering deployment and manual review.

[0006] To achieve the above objectives, this invention provides a method for extracting key elements for the automatic generation of cognitive electronic warfare test schemes, including: Obtain a dataset of cognitive electronic warfare requirements documents; Based on the aforementioned requirement document dataset, and combining task-aware template matching and location constraint priority rules, a set of candidate elements is obtained; Based on the hierarchical semantic representation, the candidate element set is semantically sorted and filtered to obtain the filtered candidate elements; The filtered candidate elements are processed according to the domain terminology knowledge base to obtain normalized elements; Based on the standardized elements, structured key element information is obtained.

[0007] Optionally, obtaining a dataset of cognitive electronic warfare requirements documents includes: Original documents were collected from the test task book, equipment technical specifications and historical test reports, and the original documents were marked with source and extracted with metadata. The original documents were subjected to double-blind annotation and arbitration according to the five-tuple annotation specification to obtain a cognitive electronic warfare requirements document dataset.

[0008] Optionally, based on the aforementioned requirement document dataset, and combining task-aware template matching and location constraint priority rules, the candidate element set is obtained, including: The dataset of required documents is preprocessed to obtain structured text; Based on the structured text, and combining task-aware template matching and position constraint priority rules, the candidate element set is obtained.

[0009] Optionally, preprocessing the aforementioned requirement document dataset to obtain structured text includes: Delete headers, footers, numbering, figure captions, and redundant explanatory content from the aforementioned requirement document dataset; The paragraphs of the demand document dataset after formatting removal are segmented into semantically independent sentences, and verb phrases and object structures are identified through dependency parsing. Based on the recognition of verb phrases and object structures, the system unifies the units of measurement, dimensions and symbolic expressions, replaces special symbolic expressions with detection limit values ​​or predefined placeholders, and replaces missing value text with unified identifiers to obtain standardized content expressions. Based on the aforementioned standard content expression, sensory descriptions and non-technical rhetorical information irrelevant to the experiment were removed.

[0010] Optionally, based on the structured text, and combining task-aware template matching and position constraint priority rules, the candidate element set is obtained as follows: Based on the structured text and combined with the task-aware template, the word matching score, distributed semantic similarity score, and contextual semantic score of the candidate elements are calculated. The matching confidence score is obtained by combining the word matching score, distributed semantic similarity score, and contextual semantic score. Based on the position index of the candidate element in the paragraph, and combined with the position constraint priority rule, the position base score is calculated by a weighted combination of exponential decay and power-law decay. Calculate the contextual consistency score between candidate elements and the sentence they belong to; The matching confidence score, location-based score, and contextual consistency score are weighted and summed to obtain a comprehensive ranking score. Candidate elements are selected based on the comprehensive ranking score to form the candidate element set.

[0011] Optionally, the candidate element set is semantically sorted and filtered according to the hierarchical semantic representation to obtain the filtered candidate elements, including: A hierarchical semantic representation based on a pre-trained language model is constructed. The hierarchical semantic representation includes word-level vectors, sentence-level vectors, and document-level vectors. Specifically, the word-level vector of each word is obtained by encoding the word segmentation sequence of the candidate elements through the pre-trained language model. Max pooling is performed on the sentence containing the word-level vector to generate the sentence-level vector, and max pooling is performed on the document containing the candidate element sentence to generate the document-level vector. Calculate the global similarity between the sentence-level vector and the document-level vector of the sentence containing the candidate element, and calculate the local similarity between the word-level embedding vector and the sentence-level vector of the candidate element; Based on the global similarity and local similarity, the semantic relevance score of the candidate elements is obtained; Construct an undirected graph among candidate elements and apply the graph centrality algorithm to calculate local importance scores; Candidate elements are sorted and filtered based on the semantic relevance score and the local importance score to obtain the filtered candidate elements.

[0012] Optionally, the filtered candidate elements are processed according to a domain terminology knowledge base to obtain normalized elements, including: Construct a domain terminology knowledge base; The filtered candidate elements are matched with knowledge base entries in the domain terminology knowledge base using character-level Jaccard similarity to obtain the knowledge matching degree; Based on the knowledge matching degree and semantic ranking results, a knowledge enhancement score is obtained; Based on the knowledge enhancement score, the candidate elements after screening in each dimension are normalized to obtain the normalized elements.

[0013] Optionally, the normalization process for candidate elements filtered by each dimension includes: Based on the knowledge enhancement score, the candidate elements after screening in each dimension are reordered, the successfully matched candidate elements are replaced with the corresponding standard terms, and the highly similar or equivalent candidate elements within the dimension are merged and deduplicated.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: This invention uses a task-aware template matching generation strategy to first identify complete term candidates, which are then confirmed and normalized by a semantic model and knowledge base, thus maintaining an advantage in the completeness of compound terms.

[0015] The hierarchical semantic modeling of this invention helps to determine the semantic boundaries of phrases in context, thereby preserving modifier information while avoiding over-segmentation.

[0016] In this invention, the knowledge base stage ensures the standardization of terminology and stores numerical information separately as attributes. This collaborative process allows the final output to be used directly by automated test plan generation, and also facilitates rapid manual review and supplementation.

[0017] This invention provides highly recall and boundary-friendly candidates in the template matching generation stage, identifies contextual relevance and preserves complete terminology in the semantic stage, and ensures terminology standardization and stores numerical information separately as attributes in the knowledge base stage. It transforms unstructured information in requirements documents into high-quality structured experimental elements that can be directly used for test plan generation and verification, thereby significantly shortening the test plan development cycle and improving the consistency and reproducibility of the plans. This collaborative process allows the final output to be used directly by automated test plan generation and also facilitates rapid manual review and supplementation. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a key element extraction method for automatically generating cognitive electronic warfare test schemes according to an embodiment of the present invention. Figure 2 This is a framework diagram of a key element extraction method for automatically generating cognitive electronic warfare test schemes according to an embodiment of the present invention; Figure 3 This is a structural diagram of the high-recall candidate generation based on the usage rules and position constraint priority rules in an embodiment of the present invention; Figure 4 This is a result diagram showing the sensitivity of the knowledge base enhancement weight parameters according to an embodiment of the present invention; Figure 5 This is a graph showing the average processing time for the key element extraction task in an embodiment of the present invention. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0021] This embodiment proposes a method for extracting key elements for the automatic generation of cognitive electronic warfare test schemes, such as... Figure 1 As shown, the specific steps include: Obtain a dataset of cognitive electronic warfare requirements documents; Based on the requirement document dataset, and combining task-aware template matching and location constraint priority rules, a set of candidate elements is obtained; Based on the hierarchical semantic representation, the candidate element set is semantically sorted and filtered to obtain the filtered candidate elements. The filtered candidate elements are processed based on the domain terminology knowledge base to obtain standardized elements; Based on standardized elements, obtain structured key element information.

[0022] Specifically, step 1) Data collection and annotation preparation. Systematically collect requirement documents in the CEW field and formulate a five-tuple annotation specification. Use a double-blind and arbitration mechanism to complete high-quality annotation, and complete the division of training / validation / test sets for subsequent model training and parameter tuning; Step 2) Text preprocessing and normalization. The collected original requirement documents are uniformly cleaned, paragraph and sentence boundaries are defined, standardized word segmentation and entity normalization are performed, special symbols and missing values ​​are handled, and irrelevant content is removed to generate normalized input. Text augmentation can be optionally performed on the training set to improve robustness to terminology variations. Step 3) Task-Aware High-Recall Candidate Generation. The high-recall candidate generation stage employs a multi-layered, progressive generation strategy, consisting of sub-modules such as text preprocessing, task-aware template matching, and positional constraint priority rules. It systematically constructs a high-coverage set of candidate elements from the CEW requirements document. The design goal of this stage is to ensure coverage as much as possible while keeping noise and redundancy within acceptable limits, providing high-quality alternatives for subsequent semantic ranking and knowledge base verification. Step 4) Hierarchical semantic representation and fine-tuning. A pre-trained Transformer is used to construct hierarchical semantic representations at the word, sentence, and document levels. The global and local relevance of candidates is calculated, and semantic scoring is performed by combining the centrality between candidates. This allows for fine-tuning of high-recall candidates to obtain element candidates with clear boundaries and high semantic relevance. Step 5) Consistency verification and terminology standardization of the knowledge base enhancement. The refined candidate terms are mapped to a flattened, multi-dimensional domain terminology knowledge base, and joint character and semantic matching, terminology replacement and deduplication are performed. When insufficient, the output is supplemented through a fallback strategy to ensure terminology standardization and output completeness.

[0023] Step 6) Evaluation and Ablation Analysis. The method is evaluated on the divided dataset using metrics such as precision, recall, and F1-score. Component ablation and parameter sensitivity analysis are performed to quantify the contribution of each module and guide the selection of hyperparameters for engineering.

[0024] As described above, this embodiment first performs structured preprocessing and task-aware template matching on the input test requirement document, generating a high-recall candidate element set based on lexical matching, distributed semantics, and positional constraint priority rules. Secondly, it constructs a hierarchical semantic representation based on a pre-trained language model, performing semantic ranking and fine-tuning of candidate elements through sentence-level and document-level global and local relevance calculations. Thirdly, it introduces a domain terminology knowledge base, employing character-level matching and smooth weighted fusion strategies to perform consistency verification, terminology standardization, and intra-dimensional re-ranking of candidate elements, and sets up a multi-layered fallback mechanism to ensure output integrity. Finally, it organizes the standardized elements into structured quintuples (equipment type, target object, functional action, capability description, and performance indicators) for test plan generation. This embodiment provides a high-coverage and standardized automatic key element extraction technology for engineering applications through the organic combination of rule-driven, semantic enhancement, and knowledge base constraints.

[0025] Furthermore, obtaining the cognitive electronic warfare requirements document dataset includes: Original documents were collected from test task books, equipment technical specifications, and historical test reports, and source annotation and metadata extraction were performed on the original documents. Double-blind annotation and arbitration were performed on the original documents according to the five-tuple annotation specification to obtain a cognitive electronic warfare requirements document dataset.

[0026] Specifically, CEW requirement documents are systematically collected from authoritative standard texts, test task books, equipment technical specifications, and historical test reports. Document metadata (such as document source, task type, equipment subdomain tags, release date, etc.) is extracted, and a five-tuple annotation specification is developed. The definitions and examples of five dimensions—"equipment type," "object of action," "functional action," "capability description," and "performance indicators"—are clearly defined to guide annotation. Double-blind annotation is employed, and inconsistent samples are arbitrated and reviewed. Cohen's calculation is performed. Consistency metrics are used to ensure annotation quality; after annotation, the data is divided into training, validation, and test sets according to a set ratio for subsequent parameter tuning and evaluation.

[0027] Furthermore, based on the requirements document dataset, and combining task-aware template matching and location constraint priority rules, the candidate element set is obtained, including: Preprocess the requirement document dataset to obtain structured text; Based on structured text, and combining task-aware template matching with location constraint priority rules, a set of candidate elements is obtained.

[0028] Furthermore, the requirement document dataset is preprocessed to obtain structured text, including: Remove headers, footers, numbering, figure captions, and redundant explanatory content from the requirements document dataset; The paragraphs of the demand document dataset after formatting removal are segmented into semantically independent sentences, and verb phrases and object structures are identified through dependency parsing. Based on the recognition of verb phrases and object structures, the system unifies the units of measurement, dimensions and symbolic expressions, replaces special symbolic expressions with detection limit values ​​or predefined placeholders, and replaces missing value text with unified identifiers to obtain standardized content expressions. Based on standardized content expression, sensory descriptions and non-technical rhetorical information irrelevant to the experiment were removed.

[0029] Specifically, text preprocessing and normalization include: removing useless information. The collected requirement documents undergo unified cleaning and standardization. First, non-textual noise such as headers, footers, figure captions, table titles, and references is removed. Paragraph and sentence boundaries are detected, and paragraph-level meta-information is established. Then, Chinese word segmentation and part-of-speech tagging are performed to identify entities such as equipment models, frequency bands, measurement methods, and units, and map them to standard word forms or placeholders. Expressions such as "<detection limit" and "not detected" are uniformly replaced with predefined placeholder strings to ensure the consistency of the input matrix format and the completeness of the values. Finally, sensory descriptions irrelevant to the experiment (such as "normal color") are removed, and synonym replacement and text enhancement are optionally performed on the training set to improve robustness to terminology variations. This preprocessing strategy serves as the basis for subsequent candidate generation and semantic encoding in the paper.

[0030] Furthermore, based on structured text, and combining task-aware template matching and positional constraint priority rules, the candidate element set is obtained, including: Based on structured text and combined with task-aware templates, the word matching score, distributed semantic similarity score, and contextual semantic score of candidate elements are calculated. The matching confidence score is obtained by combining the word matching score, the distributed semantic similarity score, and the contextual semantic score. Based on the position index of the candidate element in the paragraph, and combined with the position constraint priority rule, the position base score is calculated by a weighted combination of exponential decay and power-law decay. Calculate the contextual consistency score between candidate elements and the sentence they belong to; The overall ranking score is obtained by weighted summing of the matching confidence score, the location-based score, and the contextual consistency score. Candidate elements are then selected based on the overall ranking score to form a candidate element set.

[0031] Specifically, high-recall candidate generation is achieved using rules and positional constraints. Task-aware template matching is introduced, using predefined semantic templates to constrain text matching, thereby improving the accuracy of the candidate generation stage. Task-aware template matching employs a three-layer design, integrating lexical matching, distributed semantics, and contextual semantics signals to calculate the matching confidence score. Specifically, a lexical matching score is calculated for each candidate element. (Score based on regular expression matching), semantic similarity score based on distributed representation (Obtained through similarity after semantic clustering) and context-based semantic scores (Obtained through sentence vector cosine similarity), and then fused using linear weights to form the final matching confidence score: ; in This sets the position index of the candidate element within the paragraph. In this embodiment, the default weight setting is used. , , To fully utilize the text's positional information and contextual consistency, an adaptive positional score calculation based on self-information and contextual correction was designed. This was used to estimate the position of key elements within paragraphs based on statistical corpora. The empirical probability of " is To quantify a single location The information value is determined using pointwise surprise. Based on pointwise surprise and the primacy / recency effect, the location-based score function is defined as a weighted combination of exponential decay and power-law decay, and then normalized.

[0032] ; Among them, parameters , Control the decay rate of the exponential and power laws; , , The relative weights of the three items are: For the normalization constant, In this expression, the exponential simulation term is emphasized at the beginning; the power-law term simulates the primary effect; and the self-information term directly utilizes the "rarity" information of the expected statistics.

[0033] Then, the contextual consistency score between the candidate and the sentence is calculated. Take the cosine similarity between the candidate vector and the sentence vector and map it to... To incorporate semantic consistency, position scores are dynamically corrected using a weighted linear correction method. ; in Context correction coefficients; Finally, the candidate's overall ranking score is obtained by weighted summing of the matching confidence score and the corrected position score: ; in Control the preference for evidence matching and prior location. For all candidate keys in the paragraph... Sort the scores in descending order and take the first few. One as a candidate key element.

[0034] Furthermore, the candidate element set is semantically sorted and filtered based on the hierarchical semantic representation to obtain the filtered candidate elements, including: A hierarchical semantic representation based on a pre-trained language model is constructed. The hierarchical semantic representation includes word-level vectors, sentence-level vectors, and document-level vectors. Specifically, the word-level vector of each word is obtained by encoding the word segmentation sequence of the candidate elements through the pre-trained language model. Max pooling is performed on the sentence containing the word-level vector to generate the sentence-level vector, and max pooling is performed on the document containing the candidate element sentence to generate the document-level vector. Calculate the global similarity between the sentence-level vector and the document-level vector of the sentence containing the candidate element, and calculate the local similarity between the word-level embedding vector and the sentence-level vector of the candidate element; Based on global similarity and local similarity, semantic relevance scores of candidate elements are obtained; Construct an undirected graph among candidate elements and apply the graph centrality algorithm to calculate local importance scores; Candidate elements are sorted and filtered based on semantic relevance score and local importance score to obtain the filtered candidate elements.

[0035] Specifically, deep semantic modeling overcomes the limitations of rule-based methods in semantic understanding to quantify the relative importance of candidate phrases in a document. The encoder employs a pre-trained Transformer model, which has demonstrated strong expressive power on Chinese semantic tasks. First, the segmented sequence of the required document d is input into the BERT model to obtain the representation vectors of each token in the document. Represents token The representation vector, Document The number of tokens in the dataset. By performing max pooling on the representation vectors of the candidate elements that make up the words, we obtain the number of each candidate word. The embedding vector. Indicate candidate words The representation vector of the k-th word in the sentence is obtained. Similarly, we obtain the representation vectors of the sentence and document d respectively. and Positional deviation is measured by summing the reciprocals of each word's position in the document, and then normalized using the softmax function. Positional weights of candidate word c. : ; in, , This indicates the first occurrence position of candidate word c. The positional weights of sentence s are then assigned. This is applied to global importance calculation. Not all sentences have the same relevance to the core content of the entire document. To measure this difference, a specific function is defined based on the sentence and document representation vectors. Based on the definition of cosine similarity between the two, and combining contextual information with sentence positional weights, document-sentence relevance can be rewritten as: ; Sentence-candidate word relevance can be learned from the sentences containing the candidate words. Unlike document-sentence relevance, this metric only calculates the semantic similarity between candidate words and sentences, without considering positional bias. ; For documents Appearing in a sentence Given candidate words Document-sentence relevance for each sentence Sentence-candidate word relevance with the candidate word Multiplication, pair The final score is obtained by summing the products. .

[0036] ; To more accurately capture the importance of candidate words, we calculate the correlation (i.e., local importance) between candidate words and other candidate words. We first construct a complete undirected graph with candidate words as vertices and semantic similarity as edges; then we apply a graph centrality algorithm to calculate the importance of each candidate word. Local importance : ; in, This refers to hyperparameters determined experimentally; , which represents the mean distance between all pairs of candidate words, is used as a noise reduction factor to mitigate the impact of outliers in the document.

[0037] Obtain position weights And two scores and Then, the final score of the candidate words is calculated using a formula. Sort all candidate words in descending order of their final scores, and select the top 10 candidate words as candidate key elements: .

[0038] Furthermore, based on the domain terminology knowledge base, the filtered candidate elements are processed to obtain the standardized elements, which include: Build a domain terminology knowledge base; Character-level Jaccard similarity is used to match the filtered candidate elements with knowledge base entries in the domain terminology knowledge base to obtain the knowledge matching degree; Based on knowledge matching degree and semantic ranking results, obtain knowledge enhancement score; Based on the knowledge-enhanced score, the candidate elements after screening from each dimension are normalized to obtain normalized elements.

[0039] Furthermore, the standardization process for the candidate elements selected from each dimension includes: Based on the knowledge enhancement score, the candidate elements after filtering each dimension are reordered, the successfully matched candidate elements are replaced with the corresponding standard terms, and the highly similar or equivalent candidate elements within the dimension are merged and deduplicated.

[0040] Specifically, the knowledge base enhancement comprises five main components: constructing a terminology knowledge base covering the CEW domain, a character-level similarity-based knowledge matching mechanism, a fusion strategy of matching degree and semantic score, intra-dimensional normalization and deduplication, and multi-level fallback. Regarding knowledge base construction and organization, this embodiment constructs a flattened, multi-dimensional knowledge base including five categories of terms: standard equipment names, target objects, functional verbs, capability description templates, and performance indicator names. Knowledge sources cover authoritative standard documents, technical specifications, engineering manuals, and supplements from domain experts, while also incorporating frequency statistics from large-scale CEW documents to capture common variants. To ensure engineering maintainability, the knowledge base stores standard word forms, synonym variants, standardized classifications, and optional attributes (such as units and measurement methods) on an item-by-item basis, and establishes a regular update mechanism (such as quarterly or incremental updates as needed). To achieve robust matching of OOV and expression variants, character-level Jaccard similarity is used as the primary matching metric. Let the character set of string xxx be... ,but: ; Character-based similarity performs well in handling approximate matching of synonym variants, phrase variations, and new terms, and its computation is simple, making it suitable as a first-round fast matching metric in large-scale knowledge bases. To improve efficiency, an n-gram inverted index and hierarchical filtering strategy are employed before knowledge base matching. A higher similarity threshold is used for short strings to ensure accuracy, while a lower threshold is used for longer strings to improve recall. The threshold range and adaptive rules are configurable in the system implementation and are optimized through a validation set. The maximum knowledge matching degree of candidates is defined based on the matching results. To integrate knowledge consistency into semantic ranking, a smooth weighted fusion is adopted. ; in, This is the knowledge enhancement coefficient. This weighted form provides a controllable balance between semantic discrimination and terminology standardization, correcting obvious terminology non-standardization while avoiding misjudgment of new terms due to over-reliance on the knowledge base. After obtaining the enhancement score, each dimension is re-ranked and normalization and deduplication operations are performed. The normalization strategy is designed differently according to the characteristics of the dimension: device type focuses on retaining standard classification words and performing word segmentation and merging or main noun extraction on compound names; target object prioritizes outputting standard target categories and saving modification information as attributes; functional actions are based on standard verbs, and complex compound verbs or parallel functions can be split or labeled into multiple items; capability descriptions and performance indicators focus on name normalization and separation of unit / numerical attributes (e.g., normalizing "detection accuracy > 90%" to "detection accuracy" and recording the specific value as an attribute). After normalization, highly similar or equivalent candidates are merged, and the original variants are retained for traceability and manual review.

[0041] The final output strategy employs Top-1 selection for each dimension to ensure the determinism and usability of the results. To avoid missing items due to strict thresholds or blind spots in knowledge base coverage, a multi-layered fallback mechanism is implemented: the first layer lowers the matching threshold and retrys; the second layer falls back to the semantic ranking results of stage two; the third layer enables high-confidence alternatives from the candidate set of stage one; if no acceptable candidates are found, predefined placeholders or manual prompts are used to maintain the integrity of the output format. This fallback logic effectively ensures the integrity and engineering robustness of the quintuple output.

[0042] The following is a detailed description of this embodiment with reference to the accompanying drawings: like Figure 1 As shown, the specific steps are as follows: S11. Systematically collect authoritative and engineering-representative requirement documents in the field of cognitive electronic warfare (CEW), covering three core data sources: First, national military standards and field textbooks (such as "General Technical Requirements for Electronic Warfare Equipment" and "Technical Specifications for Cognitive Electronic Warfare Systems"), providing a theoretical basis for terminology standardization; second, technical documents from military industrial groups (including technical manuals, product specifications, and test reports from sub-fields such as radar countermeasures, communication countermeasures, and navigation countermeasures, totaling 450 documents), ensuring the engineering practicality of the corpus; and third, practical requirement documents from test bases (test task books, test outlines, and test reports from the past three years, approximately 300 documents), ensuring that the methods are adapted to real application scenarios. The original corpus contains approximately 150,000 requirement documents, with an average of 20 sentences per document.

[0043] S12. Develop a five-tuple annotation specification, clearly defining the specific definitions of equipment type (e.g., "spectrum monitoring equipment," "portable cognitive communication jamming equipment"), target (e.g., "frequency hopping communication signal"), functional actions (e.g., "real-time perception," "intelligent jamming"), capability description (e.g., "adaptive jamming capability"), and performance indicators (e.g., "individual identification accuracy"). Provide supporting annotation examples, counterexamples, and step-by-step guidelines for complex sentence structures. Employ a double-blind annotation mechanism, with three technical personnel with over three years of experience in the CEW field annotating 9,600 high-quality sentences (20-200 characters long, containing at least two CEW field terms). Calculate the annotation consistency Kappa coefficient (≥0.85). Inconsistencies are arbitrated by a senior expert with over ten years of experience in the field, ultimately resulting in 3,400 high-quality annotated samples. Divide the samples into training, validation, and test sets in an 8:1:1 ratio to ensure consistent domain distribution across datasets, which will be used for subsequent model training, parameter tuning, and performance evaluation.

[0044] S2. Standardization of the original documents includes: First, the collected CEW requirement documents are cleaned to remove non-text content and irrelevant redundant information; then, a strategy based on punctuation and semantic rules is used to divide paragraphs and sentence boundaries, using ".", "!", "!", and ";" as sentence boundary markers. Paragraph division is completed by combining the typical structure of CEW documents: "device definition—functional description—application scenario" and "capability requirements—performance parameters—constraints." Long compound sentences with more than 200 characters are split into parallel sentences using "," and ";" to ensure the semantic integrity of each sentence; then, standardized word segmentation and entity normalization are performed, using the "jieba word segmentation + CEW domain vocabulary expansion" mode to avoid mis-splitting of professional terms, unifying synonymous or near-synonymous expressions (such as "cognitive jamming equipment" and "portable cognitive communication jamming equipment") into standard terms; finally, special symbols and missing values ​​are processed, and meaningless decorative symbols ("...") are removed. Replace expressions such as "<0.5ms" and "not detected" with "detection limit 0.5ms" and "method detection limit" using symbols like "#" and "@". Mark missing term attribute information as "to be supplemented". Finally, text augmentation (optional) can be performed on the training set. Based on the CEW terminology knowledge base, synonym replacement (e.g., "intelligent interference" replaced with "adaptive interference"), sentence transformation (conversion between active and passive sentences), and key information reorganization can be performed to expand the training set samples to 4,760. Ensure that the semantics remain unchanged during the expansion process and improve the model's robustness to terminology variations.

[0045] S3. The core modules of this embodiment are as follows: Figure 2 As shown, this system systematically addresses the problem of extracting key elements in the cognitive electronic warfare field. The entire framework takes the CEW (Cognitive Electronic Warfare) requirements document as input, undergoes three progressive processing stages, and ultimately outputs a structured five-tuple of key elements containing equipment type, target, functional actions, capability description, and performance indicators. The high-recall candidate generation stage employs a multi-level, progressive generation strategy, such as... Figure 3 As shown, the system consists of sub-modules such as text preprocessing, task-aware template matching, and position constraint priority rules, systematically constructing a high-coverage candidate element set from the CEW requirements document. The design goal of this stage is to ensure coverage as much as possible while keeping noise and redundancy within an acceptable range, providing high-quality alternatives for subsequent semantic ranking and knowledge base verification.

[0046] S31. In the text preprocessing stage, to obtain contextual representations, the pre-trained language model BERT is used for word embedding to ensure a fair comparison with existing state-of-the-art models. Statistical structural analysis of the input requirements documents reveals that paragraphs often exhibit an organizational pattern of "equipment definition—functional description—application scenario" and a descriptive structure of "capability requirements—performance parameters—constraints." Based on this, paragraph-level structured representations are constructed and paragraph meta-information is output, providing statistical basis for position scores.

[0047] S32. Since key elements typically appear in semi-fixed semantic patterns, relying solely on word frequency or deep models is insufficient to accurately capture their boundaries and semantic roles with limited samples. Therefore, this embodiment introduces a task-aware template matching mechanism, which uses predefined semantic templates to constrain text matching, thereby improving the accuracy of the candidate generation stage. The task-aware template matching mechanism employs a three-layer design, integrating lexical matching, distributed semantics, and contextual semantics signals to calculate the matching confidence score. Specifically, a lexical matching score is calculated for each candidate element. (Score based on regular expression matching), semantic similarity score based on distributed representation (Obtained through similarity after semantic clustering) and context-based semantic scores (Obtained through sentence vector cosine similarity), and then fused using linear weights to form the final matching confidence score: ; in This sets the position index of the candidate element within the paragraph. In this implementation, the default weight setting is used. , , .

[0048] S33. To fully utilize the text's positional information and contextual consistency, an adaptive positional score calculation based on self-information and contextual correction was designed. This was based on estimating the position of key elements within paragraphs using statistical corpora. The empirical probability of " is To quantify a single location The information value is derived from pointwise surprise: ; The greater the self-information, the rarer the position, meaning the more informative the candidate key element, and therefore it should be given higher weight from an information value perspective. Based on self-information and the primacy / recency effect, the positional basic score function is defined as a weighted combination of exponential decay and power-law decay, and then normalized.

[0049] ; Where parameters , Control the decay rate of the exponential and power laws; , , The relative weights of the three items are: For the normalization constant, In this expression, the exponential simulation term is emphasized at the beginning; the power-law term simulates the primary effect; and the self-information term directly utilizes the "rarity" information of the expected statistics.

[0050] Then, the contextual consistency score between the candidate and the sentence is calculated. Take the cosine similarity between the candidate vector and the sentence vector and map it to... : ; To incorporate semantic consistency, position scores are dynamically corrected using a weighted linear correction method. ; in The context correction coefficients are finally obtained by weighted summing of the match confidence and the corrected position score to obtain the candidate's overall ranking score: ; in, This embodiment controls the preference for evidence matching and prior location; by default... (Emphasis on matching). For all candidate items in the paragraph... Sort the scores in descending order and take the first few. These are selected as key candidate elements. To reduce output redundancy, the selected candidates undergo post-processing in one step: identical text is deduplicated (keeping higher-order entries), and overlapping items in the text span are removed. Then merge or retain higher-scoring or longer phrases, and remove candidates that are too short or obviously noisy.

[0051] S4. Hierarchical context-aware unsupervised candidate element ranking involves constructing hierarchical semantic representations of sentences and documents using a pre-trained language model, and then measuring the global and local importance of candidate phrases. The goal of this stage is to overcome the limitations of rule-based methods in semantic understanding through deep semantic modeling, thereby quantifying the relative importance of candidate phrases in the document. The encoder uses a pre-trained Transformer model, which has shown strong expressive power on Chinese semantic tasks. To further improve the adaptability to CEW terminology through domain-adaptive fine-tuning, the specific methods are as follows: S41. Hierarchical Semantic Representation Construction: After inputting the segmented demand document sequence d into the BERT model, the representation vectors of each token in the document are obtained: ; in, Represents token The representation vector, Document The number of tokens in the system.

[0052] By performing max pooling on the representation vectors of the candidate elements that make up the words, we obtain each candidate word. Embedding vector: ; in, Indicate candidate words The representation vector of the k-th word in the sentence is obtained. Similarly, we obtain the representation vectors of the sentence and document d respectively. and .

[0053] S42. There are two main methods for quantifying positional bias: using all occurrence positions or only using the first occurrence position of the candidate word. Positional bias is measured by summing the reciprocals of the word's occurrences at each position in the document; the reciprocal of the first occurrence position is used as the positional bias weight. The latter method is adopted, and normalization is performed using the softmax function. The positional weights of the candidate words are: ; in, , Indicates the first occurrence position of candidate word c.

[0054] S43. Global and Local Relevance Calculation: Extending positional bias to the sentence level and assigning positional weights to sentence s. This is applied to global importance calculation. The hierarchical relationship between documents and sentences indicates that documents provide broader context, while sentences typically express complete ideas or semantic units. However, not all sentences are equally relevant to the core content of the entire document. To measure this difference, we define a specific function based on the representation vectors of sentences and documents: , where the function The document-sentence similarity is defined using the cosine similarity between the two sentences. Although other metrics can be used, experience shows that cosine similarity performs best. Combining contextual information with sentence position weights, document-sentence relevance can be written as: ; Sentence-candidate word relevance can be learned from the sentences containing candidate words. Unlike document-sentence relevance, this metric only calculates the semantic similarity between candidate words and sentences, without considering positional bias. The expression is: .

[0055] The candidate word mention-guided fusion method efficiently combines the two types of relevance mentioned above. Specifically, for documents... Appearing in a sentence Given candidate words Document-sentence relevance for each sentence Relevance of sentence candidate words to this candidate word Multiplication, pair The final score is obtained by summing the products. : ; To more accurately capture the importance of candidate words, S45 calculates the correlation between candidate words and other candidate words. First, a complete undirected graph is constructed, with candidate words as vertices and semantic similarity as edges; then, a graph centrality algorithm is applied to calculate the correlation between each candidate word. Local importance : ; in, This refers to hyperparameters determined experimentally; , which represents the mean distance between all pairs of candidate words, is used as a noise reduction factor to mitigate the impact of outliers in the document.

[0056] Obtain position weights And two scores and Then, the final score of the candidate words is calculated using a formula. Sort all candidate words in descending order of their final scores and select the top 10 candidate words as candidate key elements.

[0057] ; S5, the knowledge base-enhanced consistency verification and re-ranking stage, introduces domain knowledge for consistency verification and re-ranking based on semantic ranking. This aims to further map semantically driven candidates to a standardized engineering terminology system, ensuring the final output meets domain standards and experimental engineering usability. The core of this stage includes: constructing a terminology knowledge base covering the CEW domain; a character-level similarity-based knowledge matching mechanism; a fusion strategy of matching degree and semantic score; intra-dimensional normalization and deduplication; and multi-layered fallback to ensure output integrity.

[0058] S51. Joint character and semantic matching: An n-gram inverted index (n=2) and hierarchical filtering strategy are used for fast matching. First, candidate terms are retrieved by character; then, obviously mismatched terms are filtered out by length and first character; finally, character-level Jaccard similarity is calculated. ; Where A is the candidate character set, and B is the knowledge base term character set. A threshold is set for short terms (length ≤ 4 characters). Set a threshold for long terms (length > 4 characters). If the threshold is exceeded, the match is considered successful. S52. Terminology Replacement and Standardization: For successfully matched candidate elements, replace them with standard terms from the knowledge base. For example, replace "intelligent interference action" with "intelligent interference," while retaining the standard expression for "individual recognition accuracy." For capability descriptions and performance indicators, separate the name and attribute information. For example, "detection accuracy > 90%" is standardized to "detection accuracy" (indicator name) and "90%" (attribute value). Implement deduplication within the dimension, retaining the standard term with the highest score within the same dimension (such as equipment type) and deleting duplicates or synonyms.

[0059] S53. Backtracking strategy completes the output. If no matching candidate is found in a certain dimension (e.g., emerging terms not covered by the knowledge base), a three-layer backtracking strategy is initiated: The first layer lowers the matching threshold (short terms). Long terms First, rematch; then, backtrack to the semantic ranking results of S4 in the second layer and select the top candidate for that dimension; then, backtrack to the high recall candidate set of S3 in the third layer and select the candidate with the highest confidence in that dimension; if there are still no valid candidates, use a predefined placeholder tag (such as "device type - to be confirmed") to ensure that the output quintuple structure is complete.

[0060] S54. Final Structured Output: The output contains a structured quintuple of device type, target, function, capability description, and performance index. For example, if the input text is "The device under test is a portable cognitive communication jamming device that can be used to intelligently jam terrestrial frequency hopping communication signals. It has the ability to identify known radio stations individually, with an individual identification accuracy of not less than 85%", the output quintuple will be ["cognitive communication jamming device", "terrestrial frequency hopping communication signal", "intelligent jamming", "radio station individual identification capability", "individual identification accuracy"].

[0061] S6. Evaluation and Ablation Analysis. The test set partitioned by S1 is used for evaluation. Key metrics include precision, recall, and F1-score, which measure extraction accuracy, feature coverage, and overall performance, respectively. Nine representative baseline methods are selected for comparison, covering four major technical approaches: traditional statistical methods, graph ranking, pre-trained representation-based methods, and large language models. Traditional statistical methods include TF-IDF and LDA; graph ranking methods include TextRank, SingleRank, and PositionRank; pre-trained representation-based methods include SIFRank and MDERank; and large language models include Deepseek (few-shot examples) and Deepseek (domain fine-tuning), as shown in Table 1. All baseline methods and the method in this embodiment must use the same dataset and evaluation metrics to ensure fair comparison. Five experimental configurations were designed to quantify the contribution of each module: full model (HCE-KB: rules + BERT + knowledge base + fault tolerance and fallback), HCE (no knowledge base: rules + BERT, knowledge base enhancement module removed), Rule-KB (no BERT: rules + knowledge base, BERT semantic representation module removed), HCE-KB (no fault tolerance: full model with fallback strategy disabled), and Rule-Only (extraction based solely on rule templates). By comparing the Precision, Recall, and F1-score of each experimental group, the core role of each module was clarified, as shown in Table 2. Table 1 Table 2 To verify the impact of key hyperparameters on performance and ensure the stability of the method in engineering applications, a systematic sensitivity analysis was conducted on two types of key parameters: sentence vector pooling strategy and knowledge base enhancement weight α. Table 3 compares the effects of three sentence vector pooling strategies: CLS vector, MeanPooling, and MaxPooling. Figure 4 To enhance the weight of the knowledge base A systematic grid search and testing were conducted. Figure 5 The bar chart showing the average processing time of the eight methods visually illustrates the hierarchical distribution of efficiency among different methods.

[0062] Table 3 The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for key element extraction for automatic generation of cognitive electronic warfare test plan, characterized in that, include: Obtain a dataset of cognitive electronic warfare requirements documents; Based on the aforementioned requirement document dataset, and combining task-aware template matching and location constraint priority rules, a set of candidate elements is obtained; Based on the hierarchical semantic representation, the candidate element set is semantically sorted and filtered to obtain the filtered candidate elements; The filtered candidate elements are processed according to the domain terminology knowledge base to obtain normalized elements; Based on the standardized elements, structured key element information is obtained.

2. The method of claim 1, wherein the method is characterized by, The dataset of cognitive electronic warfare requirements documents includes: Original documents were collected from the test task book, equipment technical specifications and historical test reports, and the original documents were marked with source and extracted with metadata. The original documents were subjected to double-blind annotation and arbitration according to the five-tuple annotation specification to obtain a cognitive electronic warfare requirements document dataset.

3. The method for extracting key elements for automatic generation of cognitive electronic warfare test schemes according to claim 1, characterized in that, Based on the aforementioned requirement document dataset, and combining task-aware template matching and location constraint priority rules, the candidate element set is obtained as follows: The dataset of required documents is preprocessed to obtain structured text; Based on the structured text, and combining task-aware template matching and position constraint priority rules, the candidate element set is obtained.

4. The method for extracting key elements for automatic generation of cognitive electronic warfare test schemes according to claim 3, characterized in that, Preprocessing the aforementioned requirement document dataset to obtain structured text includes: Delete headers, footers, numbering, figure captions, and redundant explanatory content from the aforementioned requirement document dataset; The paragraphs of the demand document dataset after formatting removal are segmented into semantically independent sentences, and verb phrases and object structures are identified through dependency parsing. Based on the recognition of verb phrases and object structures, the system unifies the units of measurement, dimensions and symbolic expressions, replaces special symbolic expressions with detection limit values ​​or predefined placeholders, and replaces missing value text with unified identifiers to obtain standardized content expressions. Based on the aforementioned standard content expression, sensory descriptions and non-technical rhetorical information irrelevant to the experiment were removed.

5. The method for extracting key elements for automatic generation of cognitive electronic warfare test schemes according to claim 3, characterized in that, Based on the structured text, and combining task-aware template matching and position constraint priority rules, the candidate element set is obtained as follows: Based on the structured text and combined with the task-aware template, the word matching score, distributed semantic similarity score, and contextual semantic score of the candidate elements are calculated. The matching confidence score is obtained by combining the word matching score, distributed semantic similarity score, and contextual semantic score. Based on the position index of the candidate element in the paragraph, and combined with the position constraint priority rule, the position base score is calculated by a weighted combination of exponential decay and power-law decay. Calculate the contextual consistency score between candidate elements and the sentence they belong to; The matching confidence score, location-based score, and contextual consistency score are weighted and summed to obtain a comprehensive ranking score. Candidate elements are selected based on the comprehensive ranking score to form the candidate element set.

6. The method for extracting key elements for the automatic generation of cognitive electronic warfare test schemes according to claim 5, characterized in that, Based on the hierarchical semantic representation, the candidate element set is semantically sorted and filtered to obtain the filtered candidate elements, including: A hierarchical semantic representation based on a pre-trained language model is constructed. The hierarchical semantic representation includes word-level vectors, sentence-level vectors, and document-level vectors. Specifically, the word-level vector of each word is obtained by encoding the word segmentation sequence of the candidate elements through the pre-trained language model. Max pooling is performed on the sentence containing the word-level vector to generate the sentence-level vector, and max pooling is performed on the document containing the candidate element sentence to generate the document-level vector. Calculate the global similarity between the sentence-level vector and the document-level vector of the sentence containing the candidate element, and calculate the local similarity between the word-level embedding vector and the sentence-level vector of the candidate element; Based on the global similarity and local similarity, the semantic relevance score of the candidate elements is obtained; Construct an undirected graph among candidate elements and apply the graph centrality algorithm to calculate local importance scores; Candidate elements are sorted and filtered based on the semantic relevance score and the local importance score to obtain the filtered candidate elements.

7. The method for extracting key elements for automatic generation of cognitive electronic warfare test schemes according to claim 6, characterized in that, The filtered candidate elements are processed according to the domain terminology knowledge base to obtain normalized elements, including: Construct a domain terminology knowledge base; The filtered candidate elements are matched with knowledge base entries in the domain terminology knowledge base using character-level Jaccard similarity to obtain the knowledge matching degree; Based on the knowledge matching degree and semantic ranking results, a knowledge enhancement score is obtained; Based on the knowledge enhancement score, the candidate elements after screening in each dimension are normalized to obtain the normalized elements.

8. The method for extracting key elements for automatic generation of cognitive electronic warfare test schemes according to claim 7, characterized in that, The standardization process for candidate elements after filtering by each dimension includes: Based on the knowledge enhancement score, the candidate elements after screening in each dimension are reordered, the successfully matched candidate elements are replaced with the corresponding standard terms, and the highly similar or equivalent candidate elements within the dimension are merged and deduplicated.

Images