Urban underground space-oriented entity extraction method based on thinking chain

Abstract

The invention provides an entity extraction method LLM-CoT-CRF for an urban underground space based on a thinking chain. The method mainly comprises the steps of fusing a public large model and a local large model to expand urban underground space entity knowledge, and performing urban underground space entity extraction based on the large model thinking chain LLM-CoT-CRF. According to the method, a public large model obtains newest urban underground space domain knowledge through a web-search-pro search engine, in order to determine the search accuracy, a local large model judges whether obtained corpora are related to the urban underground space domain or not through a correlation thinking chain CoT, and if yes, the corpora are added into a corpus after data processing; understanding of a local large model on urban underground space related knowledge is enhanced through RAG retrieval, and meanwhile entity extraction CoT is adopted to guide a large language model to conduct entity recognition. Compared with a traditional entity extraction method which usually needs complex feature engineering, the method can automatically learn useful features by using the LLM under the guidance of the thinking chain, and the feature extraction process is simplified.

Description

Technical Field

[0001] The present invention belongs to the field of combining traditional machine learning with large models, and specifically relates to a complex multimodal spatiotemporal knowledge extraction method based on large models for underground space environmental element analysis. Background Art

[0002] 1. Information status of underground space environment

[0003] With the acceleration of urbanization, the development and utilization of underground space has become increasingly important. The complexity and diversity of the underground space environment pose many challenges to its information management. At present, the level of informationization of the underground space environment is mainly reflected in the following aspects:

[0004] Data diversity: The underground space environment involves data from multiple fields such as geology, architecture, transportation, and environment. These data are in various forms, including text, images, videos, and sensor data.

[0005] Huge amount of data: With the popularization of sensors and monitoring equipment, the amount of data on underground space environments has grown exponentially, and traditional data processing methods are difficult to cope with.

[0006] High real-time data requirements: The monitoring and management of underground space environment requires real-time data support to promptly detect and deal with potential problems.

[0007] In this context, how to efficiently acquire, process and analyze underground space environmental data has become an urgent problem to be solved.

[0008] 2. Knowledge Graph

[0009] Knowledge graph is a technology that represents knowledge in the form of a graph structure. It can effectively link multi-source heterogeneous data and provide rich semantic information. For underground space environments, building a comprehensive and accurate knowledge graph is of great significance:

[0010] Data integration and association: Through knowledge graphs, data from different sources and different types can be integrated together to reveal the associations between the data.

[0011] Knowledge representation and reasoning: Knowledge graphs can not only represent explicit knowledge, but also discover implicit knowledge through reasoning to provide support for decision-making.

[0012] Information query and visualization: Based on the knowledge graph, complex information can be queried and visualized, helping users understand the underground space environment more intuitively. Summary of the Invention

[0013] To address these issues, this paper proposes a method for extracting complex multimodal spatiotemporal knowledge based on a large model. This method augments the large model with RAG (Retrieval-Augmented Generation), customizes domain-specific thought chains, and combines them with a traditional conditional random field (CRF) large model. This method not only improves the accuracy and adaptability of entity recognition but also simplifies the feature extraction process.

[0014] A method for extracting entities from urban underground spaces based on thought chains includes the following steps:

[0015] S1, Doc-Grabbing module:

[0016] This module includes the search and processing of relevant knowledge. It issues update instructions to the public big model, which then uses the web-search-pro search tool to retrieve new corpus related to urban underground space. It then preprocesses the found urban underground space-related knowledge to obtain a preprocessed dataset (Related_Uss_corpus). Specifically, data preprocessing involves calling the document processing tool in the Related_Uss_corpus, instructing the big model to extract the main content of the article based on the CoT, count keywords and their frequency, and segment the data (by period) into tokens. These tokens are then embedded using the BERT model and stored in JSON format.

[0017] S1.1. Search for relevant corpus on urban underground space

[0018] The large model of DeepSeekV3, Zhipu Qingyan, Tongyi Qianwen, and Wenxin Yiyan uses the web-search-pro search tool to search the Internet based on keywords and return search results to complete the task of searching for urban underground space-related corpus provided to the large model;

[0019] S1.2. Main content extraction CoT

[0020] In this module, we use a combination of Zero-CoT and Few-CoT to extract the main content of articles. Zero-CoT is triggered by adding keywords (step by step), while Few-CoT guides the large model step by step based on our pre-analysis of the domain characteristics, enhancing the accuracy of main content extraction. Furthermore, by passing the theme parameter {theme} to the local large model, the model is enabled to perform segmentation based on the designed themes.

[0021] S2, Doc-Classification module:

[0022] First, we build an example of classifying documents into different fields based on their keywords, and feed it into the big model to form a chain of thinking: obtain documents - summarize document keywords - determine the corresponding fields of the keywords - and classify documents by field.

[0023] Several Doc files collected based on keywords are input into the big model, which summarizes and extracts the key words of the documents, then classifies these key words into corresponding fields, and finally outputs the classified documents belonging to each field;

[0024] S3, Doc-Correlation module:

[0025] Article relevance judgment: The large model uses the tool to call the relevance judgment CoT to judge the article relevance from three aspects. Finally, a voting algorithm is used to obtain the final result. The relevance judgment process includes: S3.1, local large model relevance judgment method, S3.2, word frequency relevance judgment method, S3.3, LDA relevance judgment method and S3.4, relevance ensemble learning.

[0026] S4, Joint Entity Extraction module:

[0027] When an entity extraction question is input, the large model calls the entity extraction CoT to extract keywords and embed them, search the private knowledge base, return the top few results with high similarity, form a prompt, and give it to the large model. The large model answers the question and extracts entities based on the context. The extracted entities are then filtered and added to our seed entity set.

[0028] S5, Entities-sift module:

[0029] This module is used to remove entities that have already been stored in the seed entity set to prevent data duplication; it also uses LLM for labeling, uses LLM's prompt to formulate the definition and background of the label, labels the identified entities, and can also remove entities that are not related to the field, thereby improving the accuracy of entity recognition.

[0030] Preferably, S3.1, local large model relevance determination method:

[0031] Judgment subject: query to obtain the main content of the document and the main content of each document in the seed corpus; method: through prompt, the main content of the document and each document in the seed corpus are inserted into the local large model. The local large model gradually analyzes the relevance of the main content of the document and each document in the seed corpus according to the relevance judgment CoT, and gives the relevance score scores = {score01, score02, ..., score0n} according to (0-1). When the maximum value max(scores) in the set scores is greater than the threshold value 0.5, it is relevant;

[0032] Large model correlation judgment CoT:

[0033] Determine the purpose of the assessment - clarify the assessment criteria - read document one - identify keywords and concepts - analyze relevance - judge relevance - read document two - identify keywords and concepts - analyze relevance - judge relevance - compare the relevance of the two documents - form a conclusion and score.

[0034] Preferably, S3.2, word frequency related judgment method:

[0035] Preliminary preparation: Remove stop words from the query documents and each document of the seed corpus to obtain the word sets of the two respectively

[0036] {doc_words} and [{seed_words01}, {seed_words02}, ..., {seed_words0n}];

[0037] Divide each document of the seed corpus into topic domains and count the top k (k=600) words with the highest frequency in all documents under the topic domain

[0038] theme={wordk:n}

[0039] Among them, wordk is a word, n is the number of occurrences, and it is used as a feature word in the subject domain; after normalizing n,

[0040]

[0041] Get the weight of each word under the subject domain,

[0042] theme={wordk:p}.

[0043] Judgment subject: query the documents and each document of the seed corpus;

[0044] Correlation judgment method:

[0045] Starting from the subject domain:

[0046] Calculate the distribution of the document's word set under each subject domain, that is, count the word sets of the document under each subject domain, and then calculate the sum of the weights of these words:

[0047]

[0048] In the formula, k is the number of word sets in the document, p n is the weight of the word in the subject domain, that is:

[0049]

[0050] Select the largest H as the subject domain of the document;

[0051] Starting from the word

[0052] Finally, BERT is used to vectorize the words {word_embedding01, word_embedding02, .., word_embedding0n} of the document word set {doc_words} and the word set [{seed_words01}, {seed_words02}, ..., {seed_words0n}] of each document under the subject domain.

[0053] [{word_embedding01, word_embedding02, .., word_embedding0n},

[0054] {word_embedding01, word_embedding02, .., word_embedding0n}],

[0055] Select the vectors of the top K words with the highest frequency of occurrence and sum them up, k = 300

[0056]

[0057]

[0058] Get vector and

[0059] By calculating the cosine similarity between the two, we can get the cosine similarity between the document and each document in the seed corpus, namely:

[0060]

[0061] doc_coses={0.1, 0.8, ..., 0.6}, when the maximum value max(doc_coses) in the set doc_coses is greater than the threshold 0.5, it is relevant.

[0062] Preferably, S3.3, LDA related judgment method:

[0063] Preliminary preparation: Use the LDA model to train each document of the seed corpus to obtain the model LDA_topic;

[0064] Judgment subject: the new documents retrieved and each document in the seed corpus;

[0065] Method: Use the model LDA_topic to obtain the topic distribution of the two topics respectively: topic = [(topic 1, p 1), (topic 2, p 2), ... (topic n, pn)], that is,

[0066] P = argmax(P(topic k|doc d))

[0067] P is the maximum probability that the document can correspond to a certain topic. The number of topics is set to k, d is the individual in the new document, and P(topic k|doc d) is the probability that document d belongs to topic k. If P is greater than the threshold (0.5), it can be determined that the document belongs to the topic.

[0068] Then compare the relevance with each document in the seed corpus under the subject domain:

[0069] Normalize the topic:

[0070]

[0071] Get the topic vector Calculate by calculating the cosine formula of the two:

[0072]

[0073] Get the cosine similarity between the document and the seed corpus lda_doc_coses = {0.1, 0.8, ..., 0.6}, when the maximum value in the set lda_doc_coses is

[0074] max=argmax({Cosine Similarity})>k

[0075] When k is a threshold of 0.5, it is considered relevant.

[0076] Preferably, S3.4, correlation ensemble learning:

[0077]

[0078] Soft voting: There are N voting paths, each path outputs the probability p that the sample belongs to each category i (j)(j=1,2,…,L), K is the total number of categories of the judgment results, w i is the weight of each voting path. If soft voting is to be used, each path needs to be given a probability of relevance rather than whether it is relevant. The weights are 0.4 for large model relevance judgment, 0.3 for word frequency relevance judgment, and 0.3 for LDA relevance judgment.

[0079] Compared with the prior art, the present invention has the following beneficial effects:

[0080] 1. The public big model obtains the latest knowledge in the field of urban underground space through the web-search-pro search engine. To ensure the accuracy of the search, the local big model uses the correlation chain of thought (CoT) to determine whether the obtained corpus is relevant to the field of urban underground space. If so, it is added to the corpus after data processing. RAG retrieval is used to enhance the local big model's understanding of relevant knowledge about urban underground space. At the same time, entity extraction CoT is used to guide the big language model for entity recognition.

[0081] 2. Compared with traditional entity extraction methods that usually require complex feature engineering, this method uses LLM under the guidance of the thinking chain to automatically learn useful features, simplify the feature extraction process, and significantly improve the efficiency of urban underground space element analysis, thereby having higher accuracy, adaptability and generalization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 This is the model diagram of the LLM-CoT-CRF of the present invention

[0083] Figure 2 is a flow chart of the present invention;

[0084] Figure 3 This is the code diagram of the Doc-Grabbing module of the present invention;

[0085] Figure 4 It is the urban underground space related corpus search map of the present invention;

[0086] Figure 5 is the Few-CoT graph of the present invention;

[0087] Figure 6 This is a system block diagram of the Doc-Classification module of the present invention;

[0088] Figure 7This is a flow chart of the Doc-Correlation module judgment of the present invention;

[0089] Figure 8 This is the CoT diagram for large model correlation judgment of the present invention;

[0090] Figure 9 This is a system block diagram of the Joint Entity Extraction module of the present invention;

[0091] Figure 10 This is a system block diagram of the Entities-sift module of the present invention. DETAILED DESCRIPTION

[0092] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0093] A method for extracting entities from urban underground spaces based on thought chains includes the following steps:

[0094] S1, Doc-Grabbing module:

[0095] This module includes the search and processing of relevant knowledge: issuing update instructions to the public big model, which obtains new corpus related to urban underground space by calling the web-search-pro search tool; then preprocessing the found urban underground space related knowledge to obtain the preprocessed data set (Related_Uss_corpus); data preprocessing specifically includes: calling the document processing tool in the tool for Related_Uss_corpus, letting the big model extract the main content of the article based on the main content, counting keywords and their frequency of occurrence, and cutting the data (cutting by period) into tokens, using the bert model to embed the tokens, and storing them in json format, such as Figure 3 As shown;

[0096] S1.1. Search for relevant corpus on urban underground space

[0097] The large model of DeepSeekV3, Zhipu Qingyan, Tongyi Qianwen and Wenxin Yiyan uses the web-search-pro search tool to search the Internet based on keywords and return the search results to complete the task of searching for urban underground space related corpus provided to the large model, such as Figure 4 As shown;

[0098] S1.2. Main content extraction CoT

[0099] In this module, we extract the main content of the article by combining Zero-CoT and Few-CoT. Zero-CoT is triggered by adding keywords (step by step), while Few-CoT guides the large model step by step based on our advance analysis of the characteristics of the field to enhance the accuracy of main content extraction. At the same time, by passing the theme parameter {theme} to the local large model, the local large model is divided according to the theme we designed. Few-CoT is as follows: Figure 5 As shown;

[0100] S2, Doc-Classification module:

[0101] First, we build an example of classifying documents into different fields based on their keywords, and feed it into the big model to form a chain of thinking: obtain documents - summarize document keywords - determine the corresponding fields of the keywords - and classify documents by field.

[0102] Several Doc files collected based on keywords are input to the big model, which summarizes and extracts the subject words of the documents, and then classifies these subject words into corresponding fields. Finally, the classified documents belonging to various fields are output, such as Figure 6 As shown;

[0103] S3, Doc-Correlation module:

[0104] Article relevance judgment: The large model uses the tool to call the relevance judgment CoT to judge the article relevance from three aspects, and finally uses the voting algorithm to obtain the final result. The relevance judgment process includes: S3.1, local large model relevance judgment method, S3.2, word frequency relevance judgment method, S3.3, LDA relevance judgment method and S3.4, relevance ensemble learning, such as Figure 7 As shown;

[0105] S4, Joint Entity Extraction module:

[0106] When the entity extraction question is input, the large model calls the entity extraction CoT, extracts keywords and embeds them, searches the private knowledge base, returns the first few results with high similarity, forms a prompt, and gives it to the large model. The large model answers the question and extracts entities based on the context; the extracted entities are filtered and finally added to our seed entity set, such as Figure 9 As shown;

[0107] S5, Entities-sift module:

[0108] This module is used to remove entities that have been stored in the seed entity set to prevent data duplication; it also uses LLM for labeling, uses LLM prompt to define the definition and background of the label, labels the identified entities, and can also remove entities that are not related to the field, thereby improving the recognition accuracy of the entity. Figure 1 、 Figure 2 and Figure 10 shown.

[0109] In this embodiment, S3.1, local large model relevance determination method:

[0110] Judgment subject: query to obtain the main content of the document and the main content of each document in the seed corpus; method: through prompt, the main content of the document and each document in the seed corpus are inserted into the local large model. The local large model gradually analyzes the relevance of the main content of the document and each document in the seed corpus according to the relevance judgment CoT, and gives the relevance score scores = {score01, score02, ..., score0n} according to (0-1). When the maximum value max(scores) in the set scores is greater than the threshold value 0.5, it is relevant;

[0111] Large model correlation judgment CoT:

[0112] Determine the purpose of the evaluation - define the evaluation criteria - read document one - identify keywords and concepts - analyze relevance - judge relevance - read document two - identify keywords and concepts - analyze relevance - judge relevance - compare the relevance of the two documents - form a conclusion and score, e.g. Figure 8 shown.

[0113] In this embodiment, S3.2, word frequency correlation determination method:

[0114] Preliminary preparation: Remove stop words from the query documents and each document of the seed corpus to obtain the word sets of the two respectively

[0115] {doc_words} and [{seed_words01}, {seed_words02}, ..., {seed_words0n}];

[0116] Divide each document of the seed corpus into topic domains and count the top k (k=600) words with the highest frequency in all documents under the topic domain

[0117] theme={wordk:n}

[0118] Among them, wordk is a word, n is the number of occurrences, and it is used as a feature word in the subject domain; after normalizing n,

[0119]

[0120] Get the weight of each word under the subject domain,

[0121] theme={wordk:p}.

[0122] Judgment subject: query the documents and each document of the seed corpus;

[0123] Correlation judgment method:

[0124] Starting from the subject domain:

[0125] Calculate the distribution of the document's word set under each subject domain, that is, count the word sets of the document under each subject domain, and then calculate the sum of the weights of these words:

[0126]

[0127] In the formula, k is the number of word sets in the document, p n is the weight of the word in the subject domain, that is:

[0128]

[0129] Select the largest H as the subject domain of the document;

[0130] Starting from the word

[0131] Finally, BERT is used to vectorize the words {word_embedding01, word_embedding02, .., word_embedding0n} of the document word set {doc_words} and the word set [{seed_words01}, {seed_words02}, ..., {seed_words0n}] of each document under the subject domain.

[0132] [{word_embedding01, word_embedding02, .., word_embedding0n},

[0133] {word_embedding01, word_embedding02, .., word_embedding0n}], select the vectors of the top K words with the highest frequency of occurrence and sum them up, k = 300

[0134]

[0135]

[0136] Get vector and

[0137] By calculating the cosine similarity between the two, we can get the cosine similarity between the document and each document in the seed corpus, namely:

[0138]

[0139] doc_coses={0.1, 0.8, ..., 0.6}, when the maximum value max(doc_coses) in the set doc_coses is greater than the threshold 0.5, it is relevant.

[0140] In this embodiment, S3.3, LDA related judgment method:

[0141] Preliminary preparation: Use the LDA model to train each document of the seed corpus to obtain the model LDA_topic;

[0142] Judgment subject: the new documents retrieved and each document in the seed corpus;

[0143] Method: Use the model LDA_topic to obtain the topic distribution of the two topics respectively: topic = [(topic 1, p 1), (topic 2, p 2), ... (topic n, pn)], that is,

[0144] P = argmax(P(topic k|doc d))

[0145] P is the maximum probability that the document can correspond to a certain topic. The number of topics is set to k, d is the individual in the new document, and P(topic k|doc d) is the probability that document d belongs to topic k. If P is greater than the threshold (0.5), it can be determined that the document belongs to the topic.

[0146] Then compare the relevance with each document in the seed corpus under the subject domain:

[0147] Normalize the topic:

[0148]

[0149] Get the topic vector Calculate by calculating the cosine formula of the two:

[0150]

[0151] Get the cosine similarity between the document and the seed corpus lda_doc_coses = {0.1, 0.8, ..., 0.6}, when the maximum value in the set lda_doc_coses is

[0152] max=argmax({Cosine Similarity})>k

[0153] When k is a threshold of 0.5, it is considered relevant.

[0154] In this embodiment, S3.4, correlation ensemble learning:

[0155]

[0156] Soft voting: There are N voting paths, each path outputs the probability p that the sample belongs to each category i (j)(j=1,2,…,K), K is the total number of categories of the judgment results, w i is the weight of each voting path. If soft voting is to be used, each path needs to be given a probability of relevance rather than whether it is relevant. The weights are 0.4 for large model relevance judgment, 0.3 for word frequency relevance judgment, and 0.3 for LDA relevance judgment.

[0157] In summary, this method utilizes the Correlation Thinking Chain (CoT) to determine whether the acquired corpus is relevant to the field of urban underground space. If so, it is added to the corpus after data processing. RAG retrieval enhances the local large model's understanding of urban underground space-related knowledge, while the Entity Extraction CoT guides the large language model for entity recognition. Compared to traditional entity extraction methods that typically require complex feature engineering, this method, guided by the CoT, utilizes the LLM to automatically learn useful features, simplifying the feature extraction process and significantly improving the efficiency of urban underground space factor analysis, thereby achieving higher accuracy, adaptability, and generalization capabilities.

[0158] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0159] The foregoing description is merely a preferred embodiment of the present invention and does not limit the present invention in any way. Although the embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for extracting entities from urban underground spaces based on thought chains, characterized in that: The method comprises the following steps: S1, Doc-Grabbing module: This module includes the search and processing of relevant knowledge. It issues update instructions to the public big model, which then uses the web-search-pro search tool to retrieve new corpus related to urban underground space. It then preprocesses the found urban underground space-related knowledge to obtain the preprocessed dataset Related_Uss_corpus. Specifically, data preprocessing involves calling the document processing tool in the Related_Uss_corpus document, instructing the big model to extract the main content of the article based on the CoT, count keywords and their frequency, and segment the data into tokens based on periods. These tokens are then embedded using the BERT model and stored in JSON format. S1.

1. Search for relevant corpus on urban underground space The large model of DeepSeekV3, Zhipu Qingyan, Tongyi Qianwen, and Wenxin Yiyan uses the web-search-pro search tool to search the Internet based on keywords and return search results to complete the task of searching for urban underground space-related corpus provided to the large model; S1.

2. Main content extraction CoT In this module, we combine Zero-CoT and Few-CoT to extract the main content of articles. Zero-CoT is triggered by adding the keyword "step by step," while Few-CoT guides the large model step by step based on our pre-analysis of the domain characteristics, enhancing the accuracy of extracting the main content. Furthermore, by passing the theme parameter {theme} to the local large model, we enable it to perform segmentation based on the themes we designed. S2, Doc-Classification module: First, we build an example of classifying documents into different fields based on their keywords, and feed it into the big model to form a chain of thinking: obtain documents - summarize document keywords - determine the corresponding fields of the keywords - and classify documents by field. Several Doc files collected based on keywords are input into the big model, which summarizes and extracts the key words of the documents, then classifies these key words into corresponding fields, and finally outputs the classified documents belonging to each field; S3, Doc-Correlation module: Article relevance judgment: The large model uses the tool to call the relevance judgment CoT to judge the article relevance from three aspects. Finally, a voting algorithm is used to obtain the final result. The relevance judgment process includes: S3.1, local large model relevance judgment method, S3.2, word frequency relevance judgment method, S3.3, LDA relevance judgment method and S3.4, relevance ensemble learning. S4, Joint Entity Extraction module: When an entity extraction question is input, the large model calls the entity extraction CoT to extract keywords and embed them, search the private knowledge base, return the top few results with high similarity, form a prompt, and give it to the large model. The large model answers the question and extracts entities based on the context. The extracted entities are then filtered and added to our seed entity set. S5, Entities-sift module: This module is used to remove entities that have already been stored in the seed entity set to prevent data duplication; it also uses LLM for labeling, uses LLM's prompt to formulate the definition and background of the label, labels the identified entities, and can also remove entities that are not related to the field, thereby improving the accuracy of entity recognition.

2. The entity extraction method for urban underground space based on thought chain according to claim 1 is characterized in that: S3.

1. Local large model relevance determination method: Judgment subject: query to obtain the main content of the document and the main content of each document in the seed corpus; method: through prompt, the main content of the document and each document in the seed corpus are inserted into the local large model. The local large model gradually analyzes the relevance of the main content of the document and each document in the seed corpus according to the relevance judgment CoT, and gives a relevance score scores = {score01, score02, ..., score0n} according to 0-1. When the maximum value max(scores) in the set scores is greater than the threshold value 0.5, it is relevant; Large model correlation judgment CoT: Determine the purpose of the assessment - clarify the assessment criteria - read document one - identify keywords and concepts - analyze relevance - judge relevance - read document two - identify keywords and concepts - analyze relevance - judge relevance - compare the relevance of the two documents - form a conclusion and score.

3. The entity extraction method for urban underground space based on thought chain according to claim 1 is characterized in that: S3.2, word frequency related judgment method: Preliminary preparation: Remove stop words from the query documents and each document of the seed corpus to obtain the word sets of the two respectively {doc_words} and [{seed_words01}, {seed_words02}, ..., {seed_words0n}]; Divide each document of the seed corpus into topic domains and count the topk (k=600) words with the highest frequency in all documents under the topic domain theme={wordk:n} Among them, wordk is a word, n is the number of occurrences, and it is used as a feature word in the subject domain; after normalizing n, Get the weight of each word under the subject domain, theme={wordk:p}. Judgment subject: query the documents and each document of the seed corpus; Correlation judgment method: Starting from the subject domain: Calculate the distribution of the document's word set under each subject domain, that is, count the word sets of the document under each subject domain, and then calculate the sum of the weights of these words: In the formula, k is the number of word sets in the document, p n is the weight of the word in the subject domain, that is: Select the largest H as the subject domain of the document; Starting from the word Finally, BERT is used to vectorize the words {word_embedding01, word_embedding02, .., word_embedding0n} of the document word set {doc_words} and the word set [{seed_words01}, {seed_words02}, ..., {seed_words0n}] of each document under the subject domain. [{word_embedding01, word_embedding02, .., word_embedding0n}, {word_embedding01, word_embedding02, .., word_embedding0n}], Select the vectors of the words with the highest frequency of occurrence in the TopK and sum them up, k = 300 Get vector and By calculating the cosine similarity between the two, we can get the cosine similarity between the document and each document in the seed corpus, namely: doc_coses={0.1, 0.8, ..., 0.6}, when the maximum value max(doc_coses) in the set doc_coses is greater than the threshold 0.5, it is relevant.

4. The entity extraction method for urban underground space based on thought chain according to claim 1 is characterized in that: S3.3, LDA related judgment method: Preliminary preparation: Use the LDA model to train each document of the seed corpus to obtain the model LDA_topic; Judgment subject: the new documents retrieved and each document in the seed corpus; Method: Use the model LDA_topic to obtain the topic distribution of the two topics respectively: topic = [(topic1, p1), (topic2, p2), ... (topic n, pn)], that is, P = argmax(P(topic k|docd)) P is the maximum probability that the document can correspond to a certain topic. The number of topics is set to k, d is the individual in the new document, and P(topic k|doc d) is the probability that document d belongs to topic k. If P is greater than the threshold (0.5), the document can be determined to belong to the topic. Then compare the relevance with each document in the seed corpus under the subject domain: Normalize the topic: Get the topic vector Calculate by calculating the cosine formula of the two: Get the cosine similarity between the document and the seed corpus lda_doc_coses = {0.1, 0.8, ..., 0.6}, when the maximum value in the set lda_doc_coses is max=argmax({Cosine Similarity})>k When k is a threshold of 0.5, it is considered relevant.

5. The entity extraction method for urban underground space based on thought chain according to claim 1 is characterized in that: S3.

4. Correlation Ensemble Learning: Soft voting: There are N voting paths, each path outputs the probability p that the sample belongs to each category i (j)(j=1,2,…,K), K is the total number of categories of the judgment results, w i is the weight of each voting path; If soft voting is used, each path needs to be given a probability of relevance rather than whether it is relevant; the weights are 0.4 for large model relevance judgment, 0.3 for word frequency relevance judgment, and 0.3 for LDA relevance judgment.

Images