Literature retrieval rule construction method for heterogeneous names of small and micro wetlands
By combining the long text semantic understanding of the BERT model and the theme interpretability of the LDA model, keywords and topics in the field of small and micro wetlands are extracted and paper search rules are generated, and the problem of limited literature search coverage of small and micro wetlands is solved, achieving a more comprehensive and accurate search effect.
Patent Information
- Application Number
- CN202510197098.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-21
- Publication Date
- 2025-06-10
AI Technical Summary
The prior art has limited coverage in the search for small and micro wetland literature, and may miss certain specific names, resulting in incomplete search rules.
The search method covering multiple disciplines is adopted to obtain experimental paper data sets, keywords and topics are extracted through BERT-LDA ensemble model, the subject influence of keywords is calculated, and paper search rules are generated through K-means clustering analysis, and the search rules set is gradually expanded and optimized.
The automated mining of heterogeneous names of small and micro wetlands has been realized, the literature search rules have been expanded, the accuracy and coverage of searches have been improved, and the cohesion and interpretability research of knowledge in the field of small and micro wetlands has been promoted.
Smart Images
Figure CN120123455A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of natural language processing, and particularly to a method for constructing a literature retrieval rule for heterogeneous names of small and micro wetlands. Background Art
[0002] Term diversity and naming heterogeneity are challenges faced in literature retrieval and information acquisition in specific fields. In-depth exploration of them helps to systematically master the domain knowledge and improve the efficiency of theoretical research and social practice. Small and micro wetlands are emerging concepts in interdisciplinary fields such as geography and ecological environment, which play multiple ecosystem service functions but have the characteristics of term diversity and naming heterogeneity. Conducting research on the term diversity of small and micro wetlands and using this as a starting point to systematically construct academic knowledge in the field is of great significance for improving the protection and management level of small and micro wetlands.
[0003] Currently, the specific names of various small and micro wetlands (such as ponds, pools, ditches, streams, spring marshes, autumn depressions, springs, artificial wetlands, etc.) are mainly obtained through manual retrieval and induction. However, the coverage of manual retrieval is limited, and some specific names may be missed, resulting in incomplete retrieval rules for small and micro wetlands. Summary of the Invention
[0004] Object of the Invention: To solve the problem of limited retrieval coverage existing in the prior art, the present invention proposes a method for constructing a literature retrieval rule for heterogeneous names of small and micro wetlands.
[0005] Technical Solution: A method for constructing a literature retrieval rule for heterogeneous names of small and micro wetlands includes the following steps:
[0006] Step 1: Adopt a retrieval method covering multiple disciplinary fields to obtain an experimental paper dataset;
[0007] Step 2: Determine the first n keywords of each paper in the experimental paper dataset, compare the first n keywords of each paper with a pre-constructed small and micro wetland term set, determine the papers related to the small and micro wetland field, and mark them to obtain a marked paper set;
[0008] Step 3: Determine the optimal number of topics of the BERT-LDA integrated model, and input the marked paper set into the BERT-LDA integrated model to obtain corresponding keywords and topics;
[0009] Step 4: Based on the keywords and topics obtained in Step 3, calculate the topic influence of each keyword, and perform K-means clustering analysis on the topic influence of each keyword to obtain multiple clusters;
[0010] Step 5: Use the keywords in the highest-scoring cluster obtained by K-means clustering analysis to generate several paper retrieval rules; apply the generated paper retrieval rules to the experimental paper dataset one by one to obtain the retrieval results; judge whether the applied paper retrieval rules meet the accuracy requirements based on the retrieval results, and only expand the satisfied paper retrieval rules into the existing paper retrieval rule set. At the same time, merge the newly appeared papers in the retrieval results into the labeled paper set to obtain a new labeled paper set; use the new labeled paper set to execute Step 3 until no new paper retrieval rules are generated, and obtain the final paper retrieval rule set;
[0011] Step 6: Use the final paper retrieval rule set to retrieve papers related to the field of small and micro wetlands.
[0012] Further, the following operations are performed in the BERT-LDA integrated model:
[0013] Use a fully connected layer to concatenate and fuse the topic probability distribution matrix extracted by the LDA model and the semantic feature matrix obtained by the BERT model in a deep learning manner and fuse them in a certain proportion to obtain a fused high-dimensional feature vector;
[0014] Perform standardization processing and dimensionality reduction operations on the fused high-dimensional feature vector in sequence to obtain the corresponding keywords and topics.
[0015] Further, the specific operations for determining the optimal number of topics of the BERT-LDA integrated model include:
[0016] Determine the value range of the number of topics of the BERT-LDA integrated model;
[0017] Take the number of topics as the number of clusters, and calculate the silhouette value under different numbers of topics one by one according to the following formula:
[0018]
[0019] where b represents the average nearest cluster distance of each sample point, and a represents the average cluster centroid clustering;
[0020] Take the number of topics with the highest silhouette value as the optimal number of topics of the BERT-LDA integrated model.
[0021] Further, in Step 4, based on the keywords and topics obtained in Step 3, calculate the topic influence of each keyword according to the following formula:
[0022]
[0023] Where TI(w|t) represents the generation probability of keyword w in topic t, p(w) is the global probability of keyword w in the entire experimental paper dataset, BERT(w) represents the BERT embedding vector of keyword w, TopicEmbedding(t) represents the embedding vector of topic t, which is calculated by the mean of the BERT embedding vectors of keywords, k is a balance parameter used to control the fusion ratio between the LDA model and the BERT model, p(w|t) represents the topic-word distribution, and Sim(·) represents the cosine similarity function.
[0024] Beneficial effects: Compared with the prior art, the present invention has the following advantages:
[0025] (1) The method of the present invention combines the long text semantic understanding advantage of the BERT model and the topic interpretability of the LDA model. From the perspective of knowledge engineering, it condenses the currently scattered domain knowledge of small and micro wetlands, promotes their conservation management, realizes the automatic mining of heterogeneous names of small and micro wetlands, expands the literature retrieval rules of small and micro wetlands, and provides a new scenario for the interpretability research of text topic mining;
[0026] (2) Through the target word vector, the method of the present invention can fully combine the context semantic information, make up for the disadvantages of the LDA topic model, train a better topic vector, and obtain a key topic recognition effect with better fine-grained and clustering accuracy. Description of the drawings
[0027] Figure 1 is a flowchart of a method for constructing a literature retrieval rule for heterogeneous names of small and micro wetlands proposed by the present invention;
[0028] Figure 2 is a flowchart of a method for constructing a literature retrieval rule for heterogeneous names of small and micro wetlands proposed by the present invention;
[0029] Figure 3 are the Cv value and Perp value of the LDA model under different numbers of topics;
[0030] Figure 4 are the silhouette values of the BERT-LDA integrated model under different numbers of topics;
[0031] Figure 5 is the visualization map of the BERT-LDA integrated model in the first round and after iteration to stability. Among them, Figure 5 in (a) is the visualization map of the BERT-LDA integrated model in the first round, Figure 5 in (b) is the visualization map of the BERT-LDA integrated model after iteration to stability;
[0032] Figure 6These are the rules for paper retrieval. Detailed implementation manners
[0033] The technical solution of this embodiment will be further elaborated below in conjunction with the accompanying drawings and embodiments.
[0034] As Figure 1 shown, this embodiment proposes a method for constructing a literature retrieval rule for heterogeneous names of small and micro wetlands, which mainly includes the following steps:
[0035] Step 1: Based on the pre-constructed term set of small and micro wetlands, mark the relevant papers in the experimental paper dataset; as Figure 2 shown, the specific operations include:
[0036] The experimental paper dataset is journal papers sourced from the Wanfang Database, with a retrieval time span from January 1, 2012, to December 31, 2022, and the retrieval was conducted in December 2023. To ensure a comprehensive and extensive acquisition of papers related to small and micro wetlands, a retrieval method covering numerous disciplinary fields was adopted. Specifically, one of the three terms "wetland", "water body", and "water area" was used as the keyword, and an "AND" connection was made with all the secondary major names in the "Catalog of Undergraduate Majors in Regular Higher Education Institutions (2023 Edition)" (an example of the retrieval rule is shown in Table 1). Chinese journal literature was selected as the research material because China's geographical environment from the frigid zone to the tropical zone includes diverse wetland types, and the Chinese context contains heterogeneous specific names of small and micro wetlands and their usage methods. The above three keywords were selected because they are basic terms in the field of water science research and can serve as a prerequisite for research related to small and micro wetlands; the disciplinary classification includes 12 major categories: philosophy, history, literature, art, education, law, economics, management, science, engineering, agriculture, and medicine, with a total of 93 first-level disciplinary classifications and 703 secondary major names. In addition, if the secondary major name contains conjunctions such as "and" or "as well as", it needs to be segmented and supplemented with words, and then an "OR" query is executed. For example, "Physical Geography and Resource Environment" is split into "Physical Geography" and "Resource Environment", and "Marine Resources and Environment" is split and supplemented to "Marine Resources" and "Marine Environment" (Table 1). To control the amount of experimental data and ensure relevance, each query is sorted according to the topic relevance, and the top 15 records are selected, resulting in 4,606 papers in the end. The text content of the papers is extracted, then deduplicated, cleaned, and segmented, and finally, the TF-IDF algorithm is used to obtain the top n keywords sorted by importance in each paper.
[0037] Table 1 Example of retrieval rules
[0038]
[0039]
[0040] The pre - constructed micro - wetland term set is obtained by summarizing representative terms and naming them from review papers in the field of micro - wetlands through expert interviews and manual reading methods, and is used as the basis for the subsequent first - round paper retrieval and marking. The above - mentioned review papers are sourced from mainstream academic conferences and journal databases.
[0041] Compare the first n keywords of each paper in the experimental paper dataset with the pre - constructed micro - wetland term set. If there is an intersection, mark this paper as belonging to the micro - wetland research field, and thus obtain the marked paper set.
[0042] Step 2: Build a BERT - LDA integrated model. As Figure 2 shown, in this embodiment, use the LDA model to extract the topic probability distribution matrix of the marked papers, and use the BERT model to extract the semantic feature matrix of the marked papers; use the fully - connected layer (Dense) in the Keras 2.1.0 development package of the Python language to concatenate and fuse the semantic feature matrix and the topic probability distribution matrix in a deep - learning manner to obtain a fused high - dimensional feature vector. Use the padding or truncation method to standardize the high - dimensional feature vector to ensure that all vector lengths are the same, and use the UMAP algorithm module in the umap 0.1.1 development package to reduce the feature vector to 2 dimensions to obtain keywords with high topic influence.
[0043] The topic probability distribution matrix provides information about the distribution of each document over different topics, and can be regarded as the semantic "label" of the document. It is the basis for understanding the topic structure and content of long texts and revealing potential knowledge. Since the topic probability features do not contain the context relationship between words, the LDA model has limitations in dealing with complex lexical dependency relationships and context information in long texts, and it is difficult to understand the diversity of word meanings and deep semantic information. For this reason, this embodiment introduces the BERT model, which uses a fine - tunable bidirectional Transformer encoder instead of a traditional recurrent neural network to achieve the superposition of self - attention, that is, the word vector can more accurately reflect the meaning in different contexts, thus effectively capturing the context semantic information.
[0044] Step 3: Determine the optimal number of topics for the BERT - LDA integrated model; the specific operations include:
[0045] First, determine the value range of the number of topics for the BERT - LDA integrated model;
[0046] Then, since the BERT - LDA integrated model is not based on probability and the number of topics is the number of clusters, calculate the silhouette value for different numbers of topics one by one to determine the optimal number of topics.
[0047] The silhouette coefficient is a typical index for explaining and verifying the consistency of data clustering, and its calculation formula is as follows:
[0048]
[0049] Among them, b represents the average nearest clustering distance of each sample point; a represents the average clustering centroid clustering, and the value range of silhouette is 0 to 1. The closer this value is to 1, the closer the two clustering sample points are to each other.
[0050] In this embodiment, the silhouette_score module in the sklearn 1.4.2 development package is used to calculate silhouette.
[0051] Step 4: Input the labeled paper set into the BERT-LDA integrated model to obtain the corresponding keywords and the number of topics with the optimal number of topics.
[0052] Step 5: Based on the keywords and topics obtained in Step 4, calculate the topic influence of each keyword, and perform K-means clustering analysis on the topic influence of each keyword to obtain multiple clusters; the specific operations include:
[0053] Input the optimal number of topics of the model into the LDA model to obtain the global word frequency p(w) and the distribution of each topic-word p(w|t), and calculate the topic influence of each keyword relative to the entire paper. In order to take into account the local topic distribution and the global word frequency characteristics and enhance the interpretability of the keywords, a linear combination method is used to calculate the topic influence of each keyword in the LDA model relative to each topic. The calculation formula is as follows:
[0054]
[0055] In the formula, TI(w|t) represents the generation probability of keyword w in topic t, p(w) is the global probability of keyword w in the entire experimental paper dataset, and k is a balance parameter with a value range of 0-1. If the probability of a keyword appearing within a topic is high (i.e., p(w|t) is high) while the global appearance probability is low (i.e., p(w) is low), its topic influence score will increase significantly.
[0056] By fusing the BERT features output by the BERT model through cosine similarity, the topic influence formula can be rewritten as:
[0057]
[0058] In the formula, BERT(w) represents the BERT embedding vector of keyword w, TopicEmbedding)t_ represents the embedding vector of topic t, which is calculated by the mean of BERT embedding of keyword w, and k controls the fusion ratio between the LDA model and the BERT model. To ensure balance, k=0.5 is taken to quantitatively calculate the topic influence of keywords extracted by the BERT-LDA model.
[0059] K-means cluster analysis was performed on the topic influence of each keyword, and the Gephi 0.10 tool was used to construct a visual map. The outer large nodes and small nodes with the same color and embedded text were used to represent the topic and keywords respectively. The connection between the two represented the topic influence, which intuitively displayed the results of BERT high-dimensional features fused with LDA features.
[0060] Step 6: For the keywords in the highest cluster obtained by K-means clustering, use the combination operation in the permutation and combination to connect 2 to more highly influential keywords to generate several paper retrieval rules; use the generated paper retrieval rules one by one to search the experimental paper data set to obtain the retrieval results; judge whether the applied paper retrieval rules meet the accuracy requirements based on the retrieval results, and only expand the satisfied paper retrieval rules to the existing paper retrieval rule set, and at the same time merge the newly appearing papers in the retrieval results into the marked paper set to obtain a new marked paper set; use the new marked paper set to execute step 3 until no new paper retrieval rules are generated, that is, the model and data are both in a stable state, and the final paper retrieval rule set is obtained.
[0061] If the first n keywords of a paper include all the keywords of the search rule, it is preliminarily determined that the paper belongs to the target field. According to the search results, it is judged whether the paper search rule used meets the accuracy requirements. The specific operation includes: based on the manual interpretation of field experts, all the papers preliminarily judged are analyzed to obtain the accuracy of the search rule. Based on the pre-set accuracy threshold, new rules that are higher than the accuracy threshold are expanded to the existing paper search rules.
[0062] Step 7: Use the final set of paper retrieval rules to search for papers related to the field of micro-wetlands.
[0063] The technical effect achieved by the construction method proposed in the embodiment is now further explained in conjunction with specific cases.
[0064] Table 2 shows the small and micro wetland term set constructed by the steps mentioned in Step 1 of the embodiment. These terms confirm professional concepts such as "small and micro wetland", "small and micro wetland group", "ecological conservation", "ecological restoration", and "natural restoration" in the "Specifications for the Protection and Management of Small and Micro Wetlands (GB / T 42481-2023)", and form a concrete naming. Compare the top 50 TF-IDF keywords of each paper in the experimental paper dataset with the small and micro wetland term set. If there is an intersection, mark this paper as belonging to the research field of small and micro wetlands, and thus obtain 298 papers related to the field.
[0065] Table 2 Small and Micro Wetland Term Set
[0066]
[0067] To illustrate that the model performance of the BERT-LDA integrated model proposed in this embodiment is superior to the LDA model, the following comparison is carried out.
[0068] Set the range of the number of topics of the LDA model and the BERT-LDA integrated model to be the same, both being 2 - 30;
[0069] Calculate the Perp value and value one by one for different numbers of topics to determine the optimal number of topics of the LDA model.
[0070] The model perplexity is a representative indicator representing the generalization ability of the model, and the calculation formula is as follows:
[0071]
[0072] Among them, logp(w d ) represents the logarithmic probability of a single document d, M represents the total number of documents, and the logarithmic probabilities of all documents are summed to obtain the logarithmic likelihood value of the entire document set, that is, the overall fitting degree; represents the total number of words in the entire document set. The perplexity Perp takes a value > 0, and the smaller the result, the stronger the generalization ability of the model.
[0073] Among them, topic coherence is an important indicator for evaluating the performance of the topic model. It judges the quality of the generated topics by measuring the semantic relevance of high-frequency words, and the calculation formula is as follows:
[0074]
[0075] In the formula, w i , w j represent two high-frequency words within the topic, P represents the set of word pairs within the topic, and Sim represents the cosine similarity algorithm that calculates the similarity between word pairs and quantifies semantic relevance. is w iand w j For the word vector representation of, the dot product · calculates the similarity of word vectors, and the denominator ||v|| is the L2 norm of the word vector, which is used for normalization. is the overall consistency score of the topic, obtained by averaging the similarity scores of all word pairs, with a value range of 0 - 1. The higher the score, the more compact the topic and the better the model performance.
[0076] Since the BERT-LDA integrated model is not based on probability and cannot directly obtain the Perp value, the optimal number of topics is determined by calculating the silhouette values for multiple numbers of clustering clusters.
[0077] Finally, by comparing the values of the optimal number of topics of the two models to evaluate their advantages and disadvantages
[0078] The Model module in the gensim 4.3.2 development package is used to calculate Perp, and the CoherenceModel module is used to calculate The silhouette_score module in the sklearn 1.4.2 development package is used to calculate silhouette.
[0079] Figure 3 、 Figure 4 respectively show the values and Perp values of the LDA model, as well as the silhouette values of the BERT-LDA integrated model, for different numbers of topics. Table 3 lists the values obtained by the two models using the optimal number of topics in the first round and after iterating to stability. Generally speaking, the values of the BERT-LDA integrated model in the first round and after iterating to stability are increased by 2.96% and 3.63% respectively compared with the LDA model. Therefore, the integrated model has better topic modeling ability. For the LDA model, the Perp values in the first round and after iterating to stability both show a trend of first slightly decreasing, then remaining stable, and then significantly increasing as the number of topics increases ( Figure 3 ). The inflection points of the slight decrease and remaining stable approximately occur when the number of topics is 3 to 5, while the inflection points of remaining stable and significantly increasing approximately occur at 17 or 18. Based on this, the effective range of the optimal number of topics of this model is initially determined to be 3 to 18. Further, when the number of topics is 4, the values of the model in the first round and after iterating to stability both reach the maximum. Considering both indicators, it is determined that the optimal number of topics for the LDA model in the first round and after iterating to stability is both 4, and its values are 0.4838 and 0.5187 respectively (Table 3); compared with other numbers of topics, the Perp value has only a slight improvement ( Figure 3), further illustrate the reliability of the four optimal topic numbers.
[0080] Table 3 Comparison of evaluations of two models
[0081]
[0082] For the BERT-LDA integrated model, to ensure balance, the fusion ratio k of the two models is set to 0.5. After the first round and after iterating to stability, the fused features achieve the highest silhouette values when the number of clusters is 4 and 3 respectively. Therefore, the optimal topic numbers of the integrated model are 4 and 3 respectively. Among them, in the first round of operation, the optimal topic number is the same as that of the LDA model, further illustrating its reliability. Input this result into the LDA model to obtain p(w) and p(w|t), and then solve the values of the integrated model in the first round and after iterating to stability, which are 0.5134 and 0.5541 respectively (Table 3).
[0083] Figure 5 The visualization maps of the BERT-LDA integrated model in the first round and after iterating to stability are shown. Figure 5 Among them, the large outer nodes and the small nodes with the same color and embedded text represent topics and keywords respectively; the connection lines between the two represent the topic influence, the thicker the line, the greater the influence, and vice versa; the red circles represent the high-influence keywords after clustering. Compared with the problem of inexplicability existing in the extraction of high-dimensional feature vector keywords in the past, in this embodiment, the 768-dimensional semantic feature vector generated by the BERT model and the 4-dimensional topic probability distribution vector generated by the LDA model are fused, and then dimensionality reduction and clustering processing are performed to obtain the visualization map of the topics, keywords and their topic influences, intuitively presenting the knowledge in the field of small and micro wetlands mined and refined from a large number of journal papers.
[0084] From the first round of operation of the model ( Figure 5 (a) in it) to the state of iterating to stability ( Figure 5 (b) in it), the number of topics of the papers is reduced from 4 to 3, the number of keywords is reduced from 87 to 72, and the average value of the topic influence is increased from 5.49×10 -3 to 5.96×10 -3, indicating that the aggregation of keywords has been enhanced. In addition, the intersection of nodes of different colors appears less frequently, indicating that the semantic space distribution characteristics of keywords are more distinct. For example, the two keywords "landscape" and "ecology" changed from belonging to two themes in the first round to belonging to the same theme after stabilization. It can be seen that after iteration, the BERT-LDA integrated model reduces the feature dimension of keywords and improves their semantic aggregation effect. Further analysis shows that the four themes mined in the first round of the model can be summarized as "urban wetland park", "wetland pollution purification", "pond aquaculture", and "wetland area change"; after iteration to stability, the category of "wetland area change" disappears, and the keywords in it are assigned to the other three themes. It is speculated that the reason is that compared with the more focused research content of other themes, the keywords in "wetland area change" (such as "area", "type", "change", etc.) mostly reflect various characteristics of small and micro wetlands from the side and play an auxiliary role in actual research, so they are classified into other categories after multiple rounds of iteration.
[0085] As Figure 5 shown, in the theme of "urban wetland park", "wetland", "city", "park", "ecology", and "landscape" are keywords with high theme influence after the model iteration to stability. In the theme of "wetland pollution purification", "constructed wetland", "plant", "microorganism", "removal", "sewage", and "system" are keywords with high theme influence after the model iteration to stability. Among them, the theme influence of "constructed wetland" is always the highest before and after iteration (>15×10 -3 ); while the theme influence of "plant" and "microorganism" increased from 11.6×10 -3 in the first round to 12.9×10 -3 and 7.1×10 -3 increased to 8.3×10 -3 . This finding corroborates previous studies that constructed wetland is the type of small and micro wetland with the largest proportion of single names in the literature (about 15%), and it is distributed in different climatic and geographical environments. In the theme of "pond aquaculture", "pond", "aquaculture", "aquatic products", and "water body" are keywords with high theme influence after the model iteration to stability.
[0086] Figure 6Shows 112 paper retrieval rules above the 70% accuracy threshold, which are respectively composed of some of the 11 keywords "wetland", "plant", "removal", "sewage", "research", "system", "microorganism", "city", "landscape", "park", "ecology". Among these retrieval rules, 109 belong to the theme of "wetland pollution purification", and the remaining 3 belong to the theme of "urban wetland park". After the first round of iteration of the BERT-LDA integrated model, 63 retrieval rules are obtained, and 61 paper tags are newly added; after the second round of iteration, no new rules are obtained, but 5 paper tags are newly added and enter a stable state, indicating that the iterative operation method enables the gradual excavation of knowledge in the field of small and micro wetlands.
[0087] Analysis Figure 6 It can be seen that the accuracy of the paper retrieval rules obtained by any 3 combinations of the keywords "wetland", "plant", "removal", "sewage", "research", "system" is between 70% and 100%. This indicates that in the experimental dataset covering multiple disciplinary fields, the papers on the purification of small and micro wetland sewage account for the vast majority, including topics such as decontamination effect and decontamination mechanism; relatively speaking, the papers on wetland plants and wetland systems are reduced, presumably because they mainly provide assistance for topics such as decontamination and purification, rather than the main research objects.
[0088] After the keyword "microorganism" is combined with the above 6 keywords, 66 new paper tags are added, and the retrieval results are highly similar to those of the paper retrieval rules where the keyword "plant" is located, indicating that there is a close relationship between plants and microorganisms in the field of small and micro wetlands.
[0089] For the rules in the theme of "urban wetland park", only the three combinations of the keywords "wetland", "city", "landscape", "park", and "wetland", "city", "landscape", "ecology", and "wetland", "city", "landscape", "park", "ecology" have a correct rate exceeding the threshold and lower than 80%. The reason is presumably that some studies on urban wetland parks focus on their more comprehensive social service value rather than a single aspect of the ecological environment, such as whether the design layout of park roads is convenient and the overall satisfaction of surrounding residents with the park. Such studies are not classified into the research category of small and micro wetlands, resulting in relatively few retrieval rules with high correct rates under this theme.
[0090] Due to space limitations, Table 4 lists the newly added typical marked papers and their associated retrieval rules. Analysis shows that the paper retrieval rules mined by the method of this embodiment, that is, the keyword combinations with high theme influence, after being connected with "wetland", "water body", and "water area" by "AND", can effectively identify relevant academic papers without the specific naming terms of small and micro wetlands, realizing the effective extension of the professional concepts in the "Specifications for the Protection and Management of Small and Micro Wetlands (GB / T 42481-2023)" and the term set obtained from the expert-reviewed review papers (Table 2). For example, Paper 2 can be retrieved by the rules 84 and 107 in Figure 6 and Paper 8 can be retrieved by the rules 30, 65, 66, 68, 73, 98, 101, 102, and 105. The above rules do not cover the existing diverse terms and names of small and micro wetlands, but through iterative paper mining and training, relevant papers in the field are automatically identified to assist in aggregating the decentralized field knowledge.
[0091] Table 4 Typical Marked Papers and Their Retrieval Rules
[0092]
[0093]
Claims
1. A method for constructing literature retrieval rules for heterogeneous names of small and micro wetlands, characterized by: The following steps are involved: Step 1: Use a search method covering multiple subject areas to obtain experimental paper datasets; Step 2: Determine the first n keywords of each paper in the experimental paper dataset, compare the first n keywords of each paper with the pre-constructed micro-wetland terminology set, determine the papers related to the micro-wetland field, mark them, and obtain the marked paper set; Step 3: Determine the optimal number of topics for the BERT-LDA ensemble model, input the labeled paper collection into the BERT-LDA ensemble model, and obtain the corresponding keywords and topics; Step 4: Based on the keywords and topics obtained in step 3, calculate the topic influence of each keyword, perform K-means clustering analysis on the topic influence of each keyword, and obtain multiple clusters; Step 5: Generate several paper retrieval rules using the keywords in the highest cluster obtained by K-means cluster analysis; use the generated paper retrieval rules one by one to search the experimental paper data set to obtain retrieval results; determine whether the applied paper retrieval rules meet the accuracy requirements based on the retrieval results, and only expand the satisfied paper retrieval rules to the existing paper retrieval rule set, and merge the newly appearing papers in the retrieval results into the marked paper set to obtain a new marked paper set; use the new marked paper set to execute step 3 until no new paper retrieval rules are generated, and obtain the final paper retrieval rule set; Step 6: Use the final set of paper retrieval rules to search for papers related to the field of micro-wetlands.
2. According to claim 1, a method for constructing literature retrieval rules for heterogeneous names of small and micro wetlands is characterized by: The following operations are performed in the BERT-LDA ensemble model: The topic probability distribution matrix extracted by the LDA model and the semantic feature matrix extracted by the BERT model are fused in series in a deep learning manner according to a certain ratio using a fully connected layer to obtain a fused high-dimensional feature vector. The fused high-dimensional feature vectors are standardized and reduced in dimension to obtain the corresponding keywords and topics.
3. According to claim 1, a method for constructing literature retrieval rules for heterogeneous names of small and micro wetlands is characterized by: The specific operations of determining the optimal number of topics for the BERT-LDA ensemble model include: Determine the range of the number of topics for the BERT-LDA ensemble model; Taking the number of topics as the number of clusters, the silhouette values under different numbers of topics are calculated one by one according to the following formula: Among them, b represents the average nearest cluster distance of each sample point, and a represents the average cluster centroid cluster; The number of topics with the highest silhouette value is taken as the optimal number of topics for the BERT-LDA ensemble model.
4. The method for constructing literature retrieval rules for heterogeneous names of small and micro wetlands according to claim 1 is characterized in that: In step 4, based on the keywords and topics obtained in step 3, the topic influence of each keyword is calculated according to the following formula: Where TI(w|t) represents the generation probability of keyword w in topic t, p(w) is the global probability of keyword w in the entire experimental paper dataset, BERT(w) represents the BERT embedding vector of keyword w, TopicEmbedding(t) represents the embedding vector of topic t, which is calculated by the mean of the BERT embedding vector of the keyword, k is a balance parameter used to control the fusion ratio between the LDA model and the BERT model, p(w|t) represents the topic-word distribution, and Sim(·) represents the cosine similarity function.