Native search type construction method and device based on large model extraction and term alignment

Abstract

The invention belongs to the technical field of computers, and particularly relates to a novelty search type construction method and device based on large model extraction and term alignment, and the method comprises the steps: constructing a term library based on historical paper iteration; extracting and acquiring different facet initial search words related to the novelty search points based on large model information; identifying the relation of the initial retrieval words based on an SPO structure to construct an initial retrieval formula, and implementing preliminary retrieval in a local paper library to obtain related papers; constructing a candidate retrieval word set based on the similarity between the initial retrieval formula and the paper sentences and the correlation between the initial retrieval words and the terms in the term library; judging whether the candidate retrieval formula and the initial retrieval formula are similar fields or not, and taking the candidate retrieval formula as a standard retrieval word when the candidate retrieval formula and the initial retrieval formula are similar fields; and combining the standard search words of different facets to obtain search formulas of different lengths. According to the method, a high-quality multi-plane novelty search formula can be automatically generated, and the efficiency and quality of scientific and technological novelty search are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of scientific and technological novelty retrieval, and specifically relates to a novelty retrieval formula construction method and device based on large model extraction and term alignment. Background Art

[0002] Scientific and technological novelty search, also known as scientific and technological literature novelty search, refers to a search and analysis activity for information resources such as patents, papers, project reports, standards and technical specifications in the field of science and technology. This process involves confirming whether a scientific and technological achievement, invention, product design or theory is novel and creative based on the retrieved literature. Scientific and technological novelty search can not only help determine whether a scientific and technological achievement is original, but also evaluate its potential application value and market prospects.

[0003] In the work of novelty search, how to construct a novelty search formula with high recall and precision is usually a difficult problem. Non-field experts often find it difficult to fully grasp professional terms, while these terms are familiar and easy to understand for field experts. In most cases, it is difficult for users to use these professional terms accurately when constructing queries. If the system can automatically supplement professional vocabulary related to the user's query intent, it will undoubtedly make it easier for users to obtain the information they need. Therefore, a method is needed to construct an accurate novelty search formula based on relatively vague novelty search points. Summary of the invention

[0004] One purpose of the present invention is to address the shortcomings of the prior art and provide a method for constructing novel search formulas based on large model extraction and term alignment. The method achieves high-quality generation of novel search formulas by constructing potential search terms and aligned terms, thereby improving the recall and precision of novel searches.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0006] A novelty search formula construction method based on large model extraction and term alignment includes the following steps:

[0007] S1, iteratively builds a term base based on a collection of historical papers;

[0008] S2, extracting information based on the big model to obtain different facet initial search terms related to the novelty point;

[0009] S3, for the initial search terms generated in step 2, identify the relationship between the initial search terms based on the SPO structure to construct an initial search formula, and conduct a preliminary search in the local paper database based on the constructed initial search formula to obtain papers related to the search point;

[0010] S4, calculate the similarity between the initial search terms and the sentences in the relevant papers obtained in step S3 to screen out the sentences that meet the requirements, extract relevant terms from the screened sentences based on the term library in S1, calculate the correlation between the initial search terms and the terms to obtain a candidate term set;

[0011] S5, using the term words in the candidate term set to replace the words in the initial search formula to obtain a candidate search formula, determining whether the candidate search formula and the initial search formula are in the same field, and selecting the replacement words in the candidate search formula that are close to the field of the initial search formula as standard search terms;

[0012] S6, based on the combination of standard search terms of different facets, obtain search expressions of different lengths.

[0013] Furthermore, the term base method in step 1 includes:

[0014] Obtain a collection of papers in related fields, obtain the abstracts of each paper from the collection, and preprocess them;

[0015] Use the big model to extract key words from the preprocessed abstract;

[0016] The TF-IDF algorithm is used to calculate the importance scores of key words, and the top N words are selected as the paper terms;

[0017] The paper and its corresponding term fields are stored in the database as the paper term base.

[0018] Furthermore, the method for obtaining the initial search term in step 2 is:

[0019] Use GPT-4's text generation capabilities to generate facet description information for each query point. The facets include research questions, research methods, and conclusions. The granularity of the description information is sentences.

[0020] Use GPT-4’s information extraction capabilities to extract facet keywords based on the facet description information above;

[0021] Finally, the keywords of each facet are deduplicated to generate an initial search term list, in which the vocabulary of the research question facet is used as the main initial search term.

[0022] Furthermore, the specific implementation method of step 3 includes:

[0023] First, the generated facet description information of the novelty query point is subjected to dependency syntactic analysis to output an SPO structure, which contains the grammatical relationship between words;

[0024] Secondly, by identifying the main initial search term and other initial search terms related to it, the logical relationship between the keywords is established to form an initial search formula;

[0025] Finally, the constructed initial search formula is used to search the local paper database to obtain specific papers related to the search point.

[0026] Furthermore, the implementation method of step 4 includes:

[0027] S41, performing sentence processing on the paper obtained in step S3 to adapt to the input format of the BERT model;

[0028] S42, loading the pre-trained BERT model, performing an embedding operation of the BERT model on each sentence in the text and the initial search formula to generate a vector representation of each sentence and the initial search formula, calculating the cosine similarity between the embedded sentence vectors, setting a first threshold, selecting sentences with a value higher than the set first threshold, and extracting relevant terms from the selected sentences;

[0029] S43. For each initial search term and related term, the cosine similarity between the initial search term and the related term is calculated using the method of S42, a second threshold is set, and terms with a value higher than the second threshold are selected to construct a candidate term set.

[0030] Furthermore, the implementation method of step 5 includes:

[0031] S51, replacing a single search term of the initial search formula with a candidate term to obtain a candidate search formula;

[0032] S52, uses historical papers as corpus to train domain classifiers;

[0033] S53, for the initial search formula and candidate search formula, use the trained domain classifier to make predictions, output the probability distribution of each category, indicating the probability that the text belongs to different fields, and then use the cross entropy loss to calculate the difference between the predicted category probabilities of the initial search formula and the candidate search formula, set the third threshold, select the candidate search formula whose difference with the initial search formula is less than the third threshold, and use the replacement words in the candidate search formula as the standard search words.

[0034] Another object of the present invention is to disclose a device for implementing the above-mentioned novelty search formula construction method based on large model extraction and term alignment, characterized in that it includes:

[0035] A term base construction module, used to iteratively construct a term base based on a collection of historical papers;

[0036] An initial search term generation module is used to obtain different facet initial search terms related to the novelty search point based on information extraction of a large model;

[0037] An initial search formula generation module, for identifying the relationship between the generated initial search terms based on the SPO structure to construct an initial search formula;

[0038] The module for obtaining relevant papers is used to conduct a preliminary search in the local paper database based on the constructed initial search formula to obtain papers related to the search point;

[0039] The candidate term set building module is used to calculate the similarity between the initial search terms and the sentences in the relevant papers to select the sentences that meet the requirements, then extract the candidate terms from the selected sentences based on the term library, and calculate the correlation between the initial search terms and the candidate terms to obtain the candidate term set;

[0040] A standard word screening module is used to replace the words in the initial search formula with the term words in the candidate term set to obtain a candidate search formula, determine whether the candidate search formula and the initial search formula are in the same field, and select the replacement words in the candidate search formula in the same field as the initial search formula as the standard search words;

[0041] The search formula construction module is used to obtain search formulas of different lengths based on the combination of standard search terms of different facets.

[0042] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0043] A computer-readable storage medium stores a computer program, which implements the steps of the above method when executed by a processor.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention constructs a terminology library based on historical paper collections, constructs an initial search formula through large model extraction, calculates the correlation between the initial search formula and the paper sentences and the terminology library, realizes the alignment of search terms and terms, and constructs a multi-dimensional professional terminology search formula, which helps to improve user experience and improve the recall rate, precision rate and efficiency of scientific and technological retrieval. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A flowchart of a novelty search formula construction method based on large model extraction and term alignment according to an embodiment of the present invention;

[0046] Figure 2 This is a specific implementation process diagram of the novelty search formula construction method based on large model extraction and term alignment according to an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present invention will be described clearly and completely below in combination with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0049] The present invention will be further described below in conjunction with specific embodiments, but the present invention is not limited thereto.

[0050] A large language model is a deep learning model with a huge number of parameters, such as GPT-3, GPT-4, etc. By learning from large-scale text data, it can understand and generate natural language, and is widely used in natural language processing tasks. In the automatic construction of retrieval formulas, the large language model can more accurately extract terms in historical papers and infer search terms related to the search points by learning from large-scale text data. By calculating the similarity between terms and search terms, the large language model can expand the search terms to build more comprehensive and high-quality search terms. This method improves the recall and precision of new searches by expanding potential search terms and aligning terms, providing researchers with a more effective information retrieval tool. Specifically, due to the diverse sources of historical data, this implementation example takes power documents as an example to illustrate the method of automatic construction of new search terms based on large model information extraction and term alignment. Figure 1 and Figure 2 As shown, the embodiment of the present invention discloses a novelty search formula construction method based on large model extraction and term alignment, comprising the following steps:

[0051] S1, iteratively builds a term base based on a collection of historical papers;

[0052] (1) First, based on a given collection of papers in the field of electricity, obtain the paper abstract and perform text preprocessing, including but not limited to filtering invalid characters, word segmentation, etc., so as to better represent the text content; in this embodiment, for example, obtain the paper abstract: In view of the low accuracy of photovoltaic power generation active power prediction, a photovoltaic power generation active power prediction algorithm based on Markov-corrected wavelet neural network is proposed. The historical data is divided into weather types such as sunny and cloudy, and trained separately to establish network models. Then, different network models are used for prediction according to the weather forecast of the forecast day. Large-scale photovoltaic grid connection will cause great disturbance to the active power balance of the power system. The dispatching department can effectively reduce the frequency fluctuation and reduce the regulation pressure of the automatic power generation control unit by using the accurate prediction of photovoltaic power generation active power to perform advance control of automatic power generation control. The actual example verifies the algorithm and simulates the frequency recovery characteristics under active power disturbance, which shows the effectiveness of the advance control method of automatic power generation control based on photovoltaic power generation active power prediction.

[0053] Among them, \u4e86 in the above text is a special character and needs to be filtered out. Therefore, it needs to be segmented. The text after segmentation is:

[0054] Aiming at the problem of low prediction accuracy of photovoltaic active power, a photovoltaic active power prediction algorithm based on Markov modified wavelet neural network is proposed. The historical data is divided into clear and cloudy weather types, and the training is carried out separately to establish the network model. Then different network models are used for prediction according to the weather forecast of the forecast day. The large-scale grid connection of photovoltaic power will cause great disturbance to the active power balance of the power system. The dispatching department can effectively reduce the frequency fluctuation and reduce the regulation pressure of the automatic generation control unit by using the accurate prediction of photovoltaic active power to carry out advance control of automatic generation control. The actual example verifies the algorithm, and simulates the frequency recovery characteristics under active disturbance, which shows the effectiveness of the advance control method of automatic generation control based on photovoltaic active power prediction.

[0055] (2) Construct a “paper keyword extraction prompt template” and use the large model to extract key words in the abstract, such as:

[0056] Input prompt template: #Task extract keywords from the paper, keywords are concatenated with;\n#Abstract: {paper abstract}\n#Keyword output:

[0057] Output keywords: photovoltaic power generation; active power prediction; Markov correction; wavelet neural network; weather type; automatic power generation control; advance control; frequency fluctuation; regulation pressure; frequency recovery characteristics.

[0058] (3) Secondly, use the TF-IDF algorithm to calculate the importance of key words in the segmented documents. The specific formula is:

[0059]

[0060]

[0061] tf-idf(w)=tf(w)*idf(w);

[0062] Among them, tf(w) refers to term frequency, which indicates the frequency of occurrence of word w in a document. n(w): the number of times word w appears in the document. N: the total number of words in the document.

[0063] idf(w): Inverse Document Frequency, used to measure whether word w is common or rare in the entire document set. Doc_N: The total number of documents in the document set. Doc_w: The number of documents containing word w. +1: Prevents division by zero errors and smoothes the IDF results.

[0064] tf-idf(w): TF-IDF value of word w, which is the product of term frequency and inverse document frequency, and is used to evaluate the importance of a word in a specific document.

[0065] Select words with high Tf-idf scores as paper terms, usually 5-10 words are retained;

[0066] The paper and its corresponding term fields are stored in the database as the paper term base.

[0067] S2, extracting information based on the big model to obtain different facet initial search terms related to the novelty point;

[0068] First, input the search point into the big model, that is, the description of a specific topic or problem, such as "Based on the constrained Markov decision model and deep reinforcement learning algorithm, a collaborative optimization control framework for active frequency is proposed to solve the problems of rapid power adjustment and optimal allocation during frequency disturbances in large power grids."

[0069] Secondly, the text generation capability of the large model is used to construct a prompt template to generate description information for each facet of the search point. The facets include research questions, research methods, and conclusions. The description information granularity is a sentence. The relevant prompt template is "#Task: Describe the research questions, research methods, and conclusions in the following abstract in one sentence respectively.\n#Abstract Based on the constrained Markov decision model and deep reinforcement learning algorithm, an active frequency collaborative optimization control framework is proposed to solve the problem of rapid power adjustment and optimal allocation during large power grid frequency disturbances.\n#Output:", the output is: "Research question: How to achieve rapid power adjustment and optimal allocation during large power grid frequency disturbances?\nResearch method: Based on the constrained Markov decision model and deep reinforcement learning algorithm, an active frequency collaborative optimization control framework is proposed.\nConclusion: The proposed control framework effectively solves the problem of rapid power adjustment and optimal allocation during large power grid frequency disturbances."

[0070] Finally, the information extraction capability of the large model is utilized to extract the key words of each facet based on the facet description information mentioned above. The input prompt template is: "#New point facet description information: From, research problem: How to achieve fast power adjustment and optimal allocation when the large power grid frequency is disturbed?\nResearch method: Based on the constrained Markov decision model and deep reinforcement learning algorithm, an active frequency collaborative optimization control framework is proposed.\nConclusion: The proposed control framework effectively solves the problem of fast power adjustment and optimal allocation when the large power grid frequency is disturbed.#Task: Extract the key words of each facet respectively:". The generated key words are as follows: "Research problem: large power grid frequency disturbance; fast power adjustment; optimal allocation\nResearch method: constrained Markov decision model; deep reinforcement learning algorithm; active frequency collaborative optimization control framework\nConclusion: control framework; large power grid frequency disturbance; fast power adjustment; optimal allocation; effective solution to ensure coverage of multiple aspects of power system reliability." Finally, the keywords of each facet are deduplicated and aggregated to generate an initial search term list, in which the words of the research problem facet are used as the main initial search terms.

[0071] S3, for the initial search terms generated in step 2, identify the relationship between the initial search terms based on the SPO structure to construct an initial search formula, and conduct a preliminary search in the local paper database based on the constructed initial search formula to obtain papers related to the search point;

[0072] Using the pyltp tool based on the SPO structure, the relationship between the initial search terms is further identified through dependency syntax analysis to construct the initial search formula. First, the generated description information is input into pyltp for dependency syntax analysis, which will output an SPO (Subject-Predicate-Object) structure, which contains the grammatical relationship between words, including but not limited to subject-predicate-object relationship, modification relationship, etc. Secondly, by identifying the main initial search term (subject) and other initial search terms related to it (predicate, object, etc.), the logical relationship between the search terms is established to form the initial search formula, in which the words with the same semantic role are represented by the or relationship, and the words with different semantic roles are represented by the and relationship, such as "large power grid frequency disturbance or rapid power adjustment" and "constrained Markov decision model and active frequency collaborative optimization control framework". Finally, in the historical paper collection, the constructed initial search formula is used to obtain specific papers related to the new search point.

[0073] S4, calculating the similarity between the initial search terms and the sentences in the relevant papers obtained in step S3 to select the sentences that meet the requirements, extracting candidate terms from the selected sentences based on the term library, and calculating the correlation between the initial search terms and the candidate terms to obtain a candidate term set;

[0074] S41, using the initial search formula constructed in S3 to preliminarily retrieve relevant documents from the collection of papers in the power field, and segment the documents into sentences to adapt to the input format of the BERT model. Load the pre-trained BERT model, and perform the BERT model embedding operation on each sentence in the document and the initial search formula to obtain the vector representation of each sentence and the initial search term to capture the semantic information. By calculating the cosine similarity between the embedded sentence vector and the initial search formula vector ( AB refers to the embedding vector, and |A||B| refers to the modulus of the quantity) to measure semantic similarity. The higher the similarity score, the stronger the semantic association between the sentences. For sentences above the set threshold, it is determined that they have a high semantic similarity with the initial search formula. Therefore, the sentences above the threshold are screened out, and the terms in the sentences are extracted from the term library built based on step 1 to build a term set;

[0075] S42, for each term in the initial search term set, use the BERT model in S41 to generate a pre-trained word vector, and calculate the cosine similarity between the word vectors of each initial search term and the term According to the calculated cosine similarity, the terms in the term set are sorted. For terms above the set threshold, it is determined that they have a high semantic relevance to the initial search terms, and other terms with low relevance to the initial search terms are filtered out to obtain a candidate term set, such as the related terms of capacity planning "capacity planning, capacity planning, capacity management, resource planning, performance planning, resource management, capacity planning, capacity management, capacity planning, capacity management, production planning, load management".

[0076] S5, using the term words in the candidate term set to replace the words in the initial search formula to obtain a candidate search formula, determining whether the candidate search formula and the initial search formula are in the same field, and selecting the replacement words in the candidate search formula in the same field as the initial search formula as the standard search terms;

[0077] S51, replacing a single search term of the initial search formula with a term in the candidate term set to obtain a candidate search formula;

[0078] S52, using historical papers as corpus to train domain classifiers: The historical paper corpus contains the label of China Library Classification Number (the most widely used classification system in domestic libraries). The training corpus is constructed using the secondary categories of China Library Classification Number, and the domain classifier is trained based on the BERT model;

[0079] S53, for the initial search formula and candidate search formula, use the trained domain classifier to make predictions, output the probability distribution of each category (indicating the probability that the text belongs to different fields), and then use the cross entropy loss (H(P,Q)=-∑(P(i)*logQ(i), where P(i) represents the probability of the i-th category in the true distribution P, and Q(i) represents the probability of the i-th category in the predicted distribution Q) to calculate the difference between the predicted category probabilities of the initial search formula and the candidate search formula, set a third threshold, select the candidate search formula whose difference with the initial search formula is less than the third threshold, and use the replacement word in the candidate search formula used to replace the initial search formula as the standard search word.

[0080] S6, based on the combination of standard search terms for different facets, different search terms of different lengths are obtained: facets include research questions, research methods, and research conclusions. Each facet may obtain one or more search terms. Search terms that do not have a determined Boolean logical relationship within a facet are connected with or, and facets are connected with and. The combined search formula contains at least one search term each for research questions and research methods, and other search terms are optional combinations.

[0081] Such as “(grid frequency disturbance or optimal allocation) and (Markov decision model or deep reinforcement learning) and (fast power adjustment or power adjustment or power allocation)”, “grid frequency disturbance and (Markov decision model or deep reinforcement learning or machine learning)”.

[0082] Another object of the present invention is to disclose a device for implementing the above-mentioned novelty search formula construction method based on large model extraction and term alignment, characterized in that it includes:

[0083] A term base construction module, used to iteratively construct a term base based on a collection of historical papers;

[0084] An initial search term generation module is used to obtain different facet initial search terms related to the novelty search point based on information extraction of a large model;

[0085] An initial search formula generation module is used to identify the relationship between the generated initial search terms based on the SPO structure to construct an initial search;

[0086] The module for obtaining relevant papers is used to conduct a preliminary search in the local paper database based on the initial search formula to obtain papers related to the search point;

[0087] The candidate term set building module is used to calculate the similarity between the initial search terms and the sentences in the relevant papers to select the sentences that meet the requirements, then extract the candidate terms from the selected sentences based on the term library, and calculate the correlation between the initial search terms and the candidate terms to obtain the candidate term set;

[0088] A standard word screening module is used to replace the words in the initial search formula with the term words in the candidate term set to obtain a candidate search formula, determine whether the candidate search formula and the initial search formula are in the same field, and select the replacement words in the candidate search formula in the same field as the initial search formula as the standard search words;

[0089] The search formula construction module is used to obtain search formulas of different lengths based on the combination of standard search terms of different facets.

[0090] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0091] A computer-readable storage medium stores a computer program, which implements the steps of the above method when executed by a processor.

[0092] The above are only preferred embodiments of the present invention, and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the contents of the specification of the present invention should be included in the protection scope of the present invention.

Claims

1. A novelty search formula construction method based on large model extraction and term alignment, characterized in that: The following steps are involved: S1, iteratively builds a term base based on a collection of historical papers; S2, extracting information based on the big model to obtain different facet initial search terms related to the novelty point; S3, for the initial search terms generated in step 2, identify the relationship between the initial search terms based on the SPO structure to construct an initial search formula, and conduct a preliminary search in the local paper database based on the constructed initial search formula to obtain papers related to the search point; S4, calculating the similarity between the initial search terms and the sentences in the relevant papers obtained in step S3 to select the sentences that meet the requirements, extracting candidate terms from the selected sentences based on the term library, and calculating the correlation between the initial search terms and the candidate terms to obtain a candidate term set; S5, replacing the words in the initial search formula with the term words in the candidate term set to obtain a candidate search formula, determining whether the candidate search formula and the initial search formula are in the same field, and selecting the replacement words in the candidate search formula in a field similar to the initial search formula as standard search terms; S6, based on the combination of standard search terms of different facets, obtain search expressions of different lengths.

2. The novelty search formula construction method based on large model extraction and term alignment according to claim 1 is characterized in that: The term base methods in step 1 include: Obtain a collection of papers in related fields, obtain the abstracts of each paper from the collection, and preprocess them; Use the big model to extract key words from the preprocessed abstract; The TF-IDF algorithm is used to calculate the importance scores of key words, and the top N words are selected as the paper terms; The paper and its corresponding term fields are stored in the database as the paper term base.

3. The novelty search formula construction method based on large model extraction and term alignment according to claim 1 is characterized in that: The method for obtaining the initial search term in step 2 is: Use GPT-4's text generation capabilities to generate facet description information for each query point. The facets include research questions, research methods, and conclusions. The granularity of the description information is sentences. Use GPT-4's information extraction capabilities to extract keywords from the above facet description information; Finally, the keywords of each facet are deduplicated and an initial search term list is constructed, in which the vocabulary of the research question facet is used as the main initial search term.

4. The novelty search formula construction method based on large model extraction and term alignment according to claim 1 is characterized in that: The specific implementation method of step 3 includes: First, the generated facet description information text is subjected to dependency syntactic analysis to output an SPO structure, wherein the SPO structure includes the grammatical relationship between words; Secondly, by identifying the main initial search terms and other initial search terms related to them, the logical relationship between the initial search terms is established to form an initial search formula; Finally, the constructed initial search formula is used to search the local paper database to obtain specific papers related to the search point.

5. The novelty search formula construction method based on large model extraction and term alignment according to claim 1 is characterized in that: Step 4 implementation methods include: S41, performing sentence processing on the paper obtained in step S3 to adapt to the input format of the BERT model; S42, loading the pre-trained BERT model, performing an embedding operation of the BERT model on each sentence in the text and the initial search formula to generate a vector representation of each sentence and each initial search formula, calculating the cosine similarity between the embedded vectors, setting a first threshold, selecting sentences higher than the set first threshold, and extracting relevant terms from the selected sentences; S43. For each initial search term and related term, the cosine similarity between the initial search term and the related term is calculated using the method of S42, a second threshold is set, and terms with a value higher than the second threshold are selected to construct a candidate term set.

6. The novelty search formula construction method based on large model extraction and term alignment according to claim 1 is characterized in that: Step 5 implementation method includes: S51, replacing a single search term of the initial search formula with a candidate term to obtain a candidate search formula; S52, uses historical papers as corpus to train domain classifiers; S53, for the initial search formula and candidate search formula, use the trained domain classifier to make predictions, output the probability distribution of each category, indicating the probability that the text belongs to different fields, and then use the cross entropy loss to calculate the difference between the predicted category probabilities of the initial search formula and the candidate search formula, set the third threshold, select the candidate search formula whose difference with the initial search formula is less than the third threshold, and use the replacement words in the candidate search formula as the standard search words.

7. A device for implementing the novelty search formula construction method based on large model extraction and term alignment as described in any one of claims 1 to 6, characterized in that: include: A term base construction module, used to iteratively construct a term base based on a collection of historical papers; An initial search term generation module is used to obtain different facet initial search terms related to the novelty search point based on information extraction of a large model; An initial search formula generation module, for identifying the relationship between the generated initial search terms based on the SPO structure to construct an initial search formula; The module for obtaining relevant papers is used to conduct a preliminary search in the local paper database based on the initial search formula to obtain papers related to the search point; The candidate term set building module is used to calculate the similarity between the initial search terms and the sentences in the relevant papers to select the sentences that meet the requirements, then extract the candidate terms from the selected sentences based on the term library, and calculate the correlation between the initial search terms and the candidate terms to obtain the candidate term set; A standard word screening module is used to replace the words in the initial search formula with the term words in the candidate term set to obtain a candidate search formula, determine whether the candidate search formula and the initial search formula are in the same field, and select the replacement words in the candidate search formula in the same field as the initial search formula as the standard search words; The search formula construction module is used to obtain search formulas of different lengths based on the combination of standard search terms of different facets.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images