A domain technology tree generation method and device based on large model knowledge reasoning

Abstract

This invention belongs to the field of natural language processing and artificial intelligence technology, and discloses a method and apparatus for generating domain technology trees based on large-scale model knowledge reasoning. The method includes: retrieving domain technology keywords from the internet and / or domain knowledge bases; obtaining and constructing instructions using candidate search content and domain technology keywords; calling a domain large-scale model to judge the relevance between the candidate search content and technology keywords in the instructions; obtaining and constructing instructions using high-value candidate search content and domain technology keywords; calling the domain large-scale model to extract domain technology triples from the instructions; and generating a domain technology tree using the extracted domain technology triples. By leveraging the powerful domain knowledge background of the domain large-scale model, in-depth analysis of key technologies and their downstream industrial chains is conducted, while simultaneously uncovering downstream and dependent technologies of key technologies. This effectively improves the completeness and accuracy of key technology context acquisition and significantly reduces the difficulty of obtaining key technology information.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and artificial intelligence, and in particular to a method and apparatus for generating domain technology trees based on large model knowledge reasoning. Background Technology

[0002] In recent years, with the development of high-tech fields such as new energy, semiconductors, and artificial intelligence, understanding the key nodes in the upstream, midstream, and downstream of these technology industry chains has become crucial for comprehending the development patterns and dynamics of the entire industry chain. Automated technology tree mining and generation within the industry chain relies on natural language processing methods. This involves analyzing massive amounts of scientific and technological text data to extract relevant technologies in the target domain and the relationships between upstream and downstream technologies. With the advancements in machine learning and deep learning, methods such as intelligent optimization algorithms, deep neural network models, and pre-trained language models have achieved leading performance in most natural language processing tasks, including text classification, named entity recognition, and relation extraction. These methods provide strong support for mining domain technology trees. For example, analyzing the quantum chip (QC) industry chain can leverage the powerful representation capabilities of pre-trained models to vectorize text from massive amounts of news and technical reports, quickly retrieving relevant information through semantic matching. Furthermore, by performing entity recognition and relation extraction on these text features, the upstream and downstream technical entities of QC can be identified, ultimately completing the construction of the industry chain technology tree.

[0003] However, there are several problems with using traditional machine learning and deep learning methods to build technology trees: First, traditional machine learning and deep learning methods rely on expert wisdom, and the technology trees built are small in scale and shallow in depth, making it difficult to meet the research needs of deep decoupling analysis in the face of constantly changing and evolving technology industry chains; Second, scientific and technological information is iterating rapidly, and traditional technology tree construction methods are unable to mine high-value target technology information in the face of massive, complex, and low-value-density information data. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technology tree construction, such as reliance on expert knowledge, small scale and shallow depth, and difficulty in mining high-value targets. It proposes a domain technology tree generation method based on large model knowledge reasoning, which realizes automated in-depth mining and correlation analysis of the hierarchical relationship of product components in the industrial chain for specific technical fields. This invention also discloses a domain technology tree generation device based on large model knowledge reasoning, providing users and research institutions with a convenient and professional method for displaying domain technology trees.

[0005] On the one hand, a method for generating domain technology trees based on large model knowledge reasoning is provided, including:

[0006] Step 1: Use domain-specific technical keywords to search the Internet and / or domain-specific knowledge bases to obtain candidate search results;

[0007] Step 2: Construct a first task instruction using the candidate search content and the domain technical keywords, and call the domain big model to judge the relevance between the candidate search content and the technical keywords in the first task instruction to obtain high-value candidate search content;

[0008] Step 3: Construct a second task instruction using the high-value candidate search content and the domain technology keywords, and call the domain big model to extract domain technology triples from the high-value candidate search content of the second task instruction. The domain technology triples take the domain technology keywords as the subject, any word with "inclusion" semantics in the relation set as the predicate, and downstream technology keywords or dependent technology keywords as the object.

[0009] Step 4: Generate a domain technology tree using the extracted domain technology triples.

[0010] An improved method for generating domain technology trees based on large-model knowledge reasoning also includes:

[0011] Step 5: When it is determined that the downstream technical keywords extracted this time need to add child nodes, the downstream technical keywords extracted this time are used as domain technical keywords, and the process jumps to step 1.

[0012] In an improved domain technology tree generation method based on large model knowledge reasoning, step 1 includes:

[0013] Step 1-1: Use domain-specific technical keywords to search the Internet and / or domain-specific knowledge bases to obtain the search results;

[0014] Step 1-2: Sort the search results from highest to lowest according to their relevance to the domain's technical keywords, and select a set number of the top-ranked search results as candidate search results;

[0015] Steps 1-3: Perform deduplication filtering and format parsing on the candidate search content to generate candidate search content in a uniform format.

[0016] In an improved domain technology tree generation method based on large model knowledge reasoning, step 2 includes:

[0017] Step 2-1: Traverse the candidate search content in the candidate search dataset, where all candidate search content constitutes the candidate search dataset;

[0018] Step 2-2: Combine candidate search results with their corresponding domain technical keywords to form key technology pairs, and use prompting engineering to form the first prompt question from the key technology pairs;

[0019] Steps 2-3: Input the first prompt question as the first task instruction into the domain big model to get the answer regarding whether the candidate search content matches the domain technical keywords;

[0020] Steps 2-4: Select candidate search terms that match the answer as high-value candidate search terms.

[0021] In an improved domain technology tree generation method based on large model knowledge reasoning, step 3 includes:

[0022] Step 3-1: Construct a set of relations, which contains predicate keywords representing hypernym-hypernym relationships;

[0023] Step 3-2: Traverse the high-value candidate search content in the high-value candidate dataset, where all high-value candidate search content constitute the high-value candidate dataset;

[0024] Step 3-3: Based on high-value candidate search content, using suggestion engineering, the second suggestion question is formed by using the domain technical keywords corresponding to the high-value candidate search content as the subject and any word with the semantic meaning of "inclusion" in the relation set as the predicate.

[0025] Steps 3-4: Use the second prompt question as the second task instruction, input the domain big model to determine whether there are predicates in the high-value candidate content that conform to the triplet technical relationship, and generate the matching domain technical triplet as the answer when the judgment result is that it exists.

[0026] On the other hand, a domain technology tree generation device based on large model knowledge reasoning is provided, including:

[0027] The data retrieval module is used to search the Internet and / or domain knowledge bases using domain technical keywords to obtain candidate search results;

[0028] The high-value content filtering module is used to construct a first task instruction using the candidate search content and the domain technical keywords, and call the domain big model to judge the relevance between the candidate search content and the technical keywords in the first task instruction to obtain high-value candidate search content.

[0029] The technology triple extraction module is used to construct a second task instruction using the high-value candidate search content and the domain technology keywords, and to call the domain large model to extract domain technology triples from the high-value candidate search content of the second task instruction. Each domain technology triple has a domain technology keyword as the subject, any word with "inclusion" semantics within the relation set as the predicate, and downstream technology keywords or dependent technology keywords as the object.

[0030] The domain technology tree generation module is used to generate a domain technology tree using the extracted domain technology triples.

[0031] An improved domain technology tree generation device based on large model knowledge reasoning also includes:

[0032] The domain technology tree display module is used to trigger the data retrieval module when it is determined that the extracted downstream technology keywords need to be added to the domain technology keywords.

[0033] In an improved domain technology tree generation device based on large model knowledge reasoning, the data retrieval module includes: a retrieval unit, a sorting unit, and a filtering unit.

[0034] The retrieval unit is used to retrieve the Internet and / or domain knowledge bases using domain technical keywords to obtain the retrieval content;

[0035] The sorting unit is used to sort the search content from high to low according to the relevance of the search content to the domain technical keywords, and take the first set number of search content as candidate search content.

[0036] The filtering unit is used to perform deduplication filtering and format parsing on candidate search content to generate candidate search content in a uniform format.

[0037] In an improved domain technology tree generation device based on large model knowledge reasoning, the high-value content filtering module includes: a traversal unit, a first prompt question composition unit, a matching unit, and a high-value content construction unit;

[0038] The traversal unit is used to traverse the candidate search content in the candidate search dataset, wherein all candidate search content constitutes the candidate search dataset.

[0039] The first prompt question unit is used to combine candidate search content with its corresponding domain technical keywords to form key technology pairs, and use prompt engineering to form the key technology pairs into the first prompt question;

[0040] The matching unit is used to input the first prompt question as the first task instruction into the domain big model and obtain an answer regarding whether the candidate search content matches the domain technical keywords.

[0041] The high-value content construction unit is used to select candidate search content whose answers match as high-value candidate search content.

[0042] In an improved domain technology tree generation device based on large model knowledge reasoning, the technology triple extraction module includes: a relation set construction unit, a traversal unit, a second prompt question composition unit, and a technology triple generation unit;

[0043] The relation set construction unit is used to construct a relation set, wherein the relation set contains predicate keywords representing hierarchical relationships;

[0044] The traversal unit is used to traverse the high-value candidate retrieval content in the high-value candidate dataset, wherein all high-value candidate retrieval content constitutes the high-value candidate dataset.

[0045] The second prompt question component unit is used to form a second prompt question based on high-value candidate search content, using prompt engineering, with the domain technical keywords corresponding to the high-value candidate search content as the subject and any word with the semantic meaning of "inclusion" in the relation set as the predicate;

[0046] The technical triple generation unit is used to input the second prompt question as a second task instruction into the domain big model to determine whether there is a predicate in the high-value candidate content that conforms to the technical relationship of the triple. When the determination result is that there is, a matching domain technical triple is generated as the answer.

[0047] Compared with the prior art, the advantages of the present invention are:

[0048] 1. This invention provides a method for obtaining high-quality retrieval information. By combining knowledge base retrieval, keyword matching, semantic similarity calculation, and relevance ranking, it achieves simple and efficient domain-specific text retrieval. Furthermore, by incorporating cross-domain knowledge background from a large domain model, the retrieval content is further evaluated and filtered, ensuring the acquisition of high-value retrieval content matching the keywords.

[0049] 2. This invention addresses the problem of decoupling key and difficult points in domain technologies in traditional technology tree generation by employing extracted technology triples as the primary basis for technology tree generation, thus providing a more robust technology tree generation method. Simultaneously, leveraging the powerful domain knowledge background of the large domain model, it deeply mines key technologies and their downstream industrial chains, while also uncovering downstream and dependent technologies of key technologies. This effectively improves the completeness and accuracy of key technology context acquisition, significantly reducing the difficulty of obtaining key technology information.

[0050] 3. The solution in this embodiment of the invention utilizes a large domain model with strong domain background knowledge. Furthermore, it first extracts high-value candidate search content from the candidate search content, and then extracts technical triples from these high-value candidate search content. This effectively eliminates non-high-value candidate search content, which constitutes a large portion of the overall data, from the outset. Compared to models like BERT used for small tasks, the approach of first extracting technical triples from candidate search content and then performing denoising and other normalization operations on these technical triples significantly reduces the amount of data processing required for triple extraction. Moreover, because the large domain model relies on large parameters and rich background knowledge, it is more accurate in selecting high-value candidate search content, resulting in higher-quality search content. This helps improve the quality of technical triple extraction from the data source, thereby improving the quality of the generated domain data tree. Attached Figure Description

[0051] Figure 1 This is one of the flowcharts of the domain technology tree generation method based on large model knowledge reasoning in this invention.

[0052] Figure 2 This is the second flowchart of the domain technology tree generation method based on large model knowledge reasoning in this invention.

[0053] Figure 3 This is a key technology for retrieving data in an embodiment of the present invention;

[0054] Figure 4 This is a high-value retrieval content and technology triple extraction example from an embodiment of the present invention;

[0055] Figure 5 This is a schematic diagram of the domain technology tree generation device based on large model knowledge reasoning according to an embodiment of the present invention. Detailed Implementation

[0056] In the embodiment of the present invention, a first prompt question and a second prompt question are constructed by using domain technology keywords and two search contents. The same domain big model with strong domain background knowledge is called twice to generate high-value candidate search contents and extract domain technology triples, so that it can mine high-quality domain technology trees from the search contents to meet the growing demand for domain technology tree construction.

[0057] The domain-specific large model in this embodiment of the invention is a model obtained by fine-tuning a trained large model using a specific domain-specific knowledge base. For example, a domain-specific large model can be obtained by fine-tuning an already trained open-source large model such as Tongyi Qianwen, Baichuan, or ChatGLM3 using a corpus containing knowledge from multiple aspects such as the quantum domain. In this case, the solution of this embodiment of the invention can use the domain-specific large model to construct a domain technology tree that includes cutting-edge technologies such as the quantum domain technology tree.

[0058] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0059] like Figure 1 The diagram shown is a flowchart of a domain technology tree generation method based on large model knowledge reasoning provided in an embodiment of the present invention, which includes the following steps:

[0060] Step 1: Obtain the domain technology keywords needed to build the domain technology tree, use the domain technology keywords to search the Internet and / or domain knowledge bases, perform data preprocessing operations such as format constraints on the search content, and obtain candidate search content;

[0061] Specifically, step 1 may include steps 1-1 to 1-3:

[0062] Step 1-1: Utilize domain-specific technical keywords q i Search the internet and / or domain-specific knowledge bases to obtain the search results. i ;

[0063] The search results obtained here can be paragraphs or text fragments.

[0064] The retrieval process involves identifying and acquiring information system resources relevant to information needs, given a knowledge base. Each knowledge keyword k i Corresponding to knowledge text v i .

[0065] Taking quantum chips as an example, by searching for "components of quantum chips" and "technology relied upon by quantum chips," several text fragments containing the term "quantum chip" were obtained. i For example, “a cooperation program launched by the EU and its member states at the end of 2018…in which core projects in the fields of quantum technology and human brain mapping include the research and application of quantum chips…”.

[0066] Step 1-2: Search for content v i Reorder based on technical keywords q i Candidate search terms are obtained by ranking relevance from high to low. i .

[0067] Given the technical keywords q to be queried i The similarity function s is used to search for the k most similar keys and obtain the text content v′ of the paired values. i .

[0068] Steps 1-3: Analyze candidate search terms v′ i Perform deduplication filtering and format parsing to generate candidate search content v′ in a uniform format. i .

[0069] Following the previous example, remove search segments that do not include "quantum chip" and remove candidate search segments that contain illegal characters.

[0070] In steps 1-3, removing text information that does not contain technical keywords and duplicate content from the search results transforms the candidate data into a easily processed list format, ultimately yielding the candidate search data set D. cand .like Figure 3 The example shown shows that the technical keyword "extreme ultraviolet lithography machine" was searched. After searching the knowledge base and filtering and sorting, the final candidate search data set is as follows: ["Capillary discharge electrode ultraviolet light source extreme ultraviolet lithography is considered the preferred technology for the next generation of lithography and will replace optical lithography as the mainstream lithography technology for semiconductor production after 2015. The extreme ultraviolet lithography system is a comprehensive high-tech...", "X. Breakthrough in ultraviolet lithography technology In February 2020, the Belgian Microelectronics Research Centre and ASML of the Netherlands used the NXE:3400B extreme ultraviolet lithography equipment..."].

[0071] Step 2: Construct a first task instruction using the candidate search content and the domain technical keywords, and call the domain big model to judge the relevance between the candidate search content and the technical keywords in the first task instruction to obtain high-value candidate search content.

[0072] In step 2, the retrieved content and keywords are combined to construct the first task instruction. The domain-wide model is then used to determine the relevance of this instruction, filtering and retaining high-value search information. Here, the first task instruction serves as the input judgment instruction for the large model. Subsequently, the domain-wide model is used to judge each task instruction, determining whether the current search content is relevant to the given query's technical keywords. If relevant, it is considered high-value candidate search content.

[0073] Specifically, step 2 may include steps 2-1 to 2-4:

[0074] Step 2-1: Traverse the dataset D consisting of all candidate search terms. cand Candidate search content v′ i ;

[0075] Step 2-2: Change v′i Its corresponding technical keyword q i Composition of key technologies <v′ i ,q i >, using the prompting process to form the first prompting question S i ;

[0076] like Figure 4 As shown, the first prompt question constructed from the technical keyword "extreme ultraviolet lithography machine" and the search content "X. Breakthrough in ultraviolet lithography technology: In February 2020, the Belgian Microelectronics Research Centre and ASML of the Netherlands used the NXE:3400B extreme ultraviolet lithography equipment..." could be: "Assuming you are a very good industry chain analyst, please determine whether the following candidate statements contain content related to downstream technologies of quantum chips. If yes, answer True; if not, answer False and provide an explanation."

[0077] Steps 2-3: Address the first prompt question S i As the first task instruction, the domain big model is invoked to process the data using the following formula to obtain information about the candidate search content v′. i Technical keywords q i Does the answer match?

[0078] Specifically, a sampling decoding method can be used to call the large model to generate the corresponding decision answer. Based on the autoregressive decoding method of the large model, the decision answer is obtained according to the following formula.

[0079]

[0080] Where P is the probability of generating the next character, and S... i This is a vector of all the text characters used to prompt the question. T is the length limit generated by the domain-wide model, w is the currently generated character during the model's generation process, I is the index of the currently generated word, and t is a random variable ranging from 1 to the maximum number of words generated in the sentence. Steps 2-4: Select the answer as a matching candidate search content. As high-value candidate search content.

[0081] Here, the dataset consisting of all high-value candidate search results can be used as the high-value candidate dataset D. hv .

[0082] Leveraging the powerful knowledge base and natural language reasoning capabilities of large-scale models, we can determine whether the currently retrieved information is relevant to key technologies, thereby filtering out high-value candidate datasets D that are closely related to these key technologies. hv .

[0083] Using the example above, "The text above mentions the main components of a 'quantum chip,' so we return True, and the explanation is as follows: A quantum computer is a computer based on the principles of quantum mechanics. Its basic unit is the qubit. Unlike the bits of a traditional computer, a qubit can be in multiple states simultaneously, which makes quantum computers more efficient and have stronger computing power than traditional computers when dealing with certain specific problems."

[0084] Step 3: Construct a second task instruction using the high-value candidate search content and the domain technology keywords, and call the domain big model to extract domain technology triples from the high-value candidate search content of the second task instruction.

[0085] In this context, the domain technology triplet takes the domain technology keyword as the subject, any word with the semantic meaning of "inclusion" within the relation set as the predicate, and the downstream technology keyword or dependent technology keyword as the object.

[0086] In step 3, for the filtered high-value information, instructions are constructed by combining technical keywords and other prompts. The domain big model is called to extract key technology triples, that is, to mine the hierarchical nodes and dependency nodes of key technologies. The extracted results can be preprocessed to filter out unreasonable triples.

[0087] Specifically, step 3 may include steps 3-1 to 3-4:

[0088] Step 3-1: Construct a set of relations, which contains predicate keywords representing hypernym-hypernym relationships;

[0089] Step 3-2: Traverse the high-value candidate dataset D hv High-value candidate search content;

[0090] Step 3-3: Based on high-value candidate search content, using suggestion engineering, a second suggestion question is formed, with the domain technical keywords corresponding to the high-value candidate search content as the subject and any word with the semantic meaning of "inclusion" within the relation set as the predicate.

[0091] Using the example of the two-word chip above, the second prompt question generated here... It could be: "Assuming you are a highly skilled industry chain analyst, please extract information about downstream technologies of quantum chips from the given text and return it in JSON format as a triple, with the triple format being {"head": quantum chip,"relation": constituent components,"tail": downstream technology}.

[0092] Steps 3-4: Prompt the second question It serves as the second task instruction, invoking the domain-wide model to identify high-value candidate content. If a predicate that conforms to the technical relationship of a triple exists, and the result is yes, a matching technical triple is generated as the answer.

[0093] Using the quantum chip example, the dependent technology of the quantum chip is successfully extracted here as the dependent technology triple of the control electronic device, which is "{'head':'quantum chip','relation':'dependent on','tail':'control electronic device'}".

[0094] like Figure 4 The example shown illustrates this. Given the technical keyword "extreme ultraviolet lithography machine" and one of its corresponding high-quality search results, "X. Breakthrough in Ultraviolet Lithography Technology: In February 2020, the Belgian Microelectronics Research Centre and ASML of the Netherlands utilized the NXE:3400B extreme ultraviolet lithography equipment…", The content is: "Extract the relational triples about the components of the extreme ultraviolet lithography machine from the following text and answer in the format of JSON string [{'head':”,'relation':”,'tail':”},…]\n{input text}". After processing by the large model, an answer is generated: "The main components of the extreme ultraviolet lithography machine are mentioned in the above text, therefore return True". Specifically, the text mentions the following key components: 1. Extreme Ultraviolet Lloyd's Ltd. (EUVLLC)... Meanwhile, the natural language responses generated by the large model are parsed to remove those that do not conform to the JSON format and those that do not conform to the given subject-verb agreement. The final technical triplet is: [{'head':'Extreme Ultraviolet Lithography Machine','relation':'Components','tail':'EUV Light Source'},{'head':'Extreme Ultraviolet Lithography Machine','relation':'Components','tail':'Illumination Optical System'},{'head':'Extreme Ultraviolet Lithography Machine','relation':'Components','tail':'Miniature Projection Optical System'}...].

[0095] Step 4: Construct a technology tree using the extracted technology triples.

[0096] Step 5: When it is determined that the downstream technical keywords extracted this time need to add child nodes, the downstream technical keywords extracted this time are used as domain technical keywords, and the process jumps to step 1.

[0097] Specifically, by inputting key technologies, such as quantum chips, the method will complete steps 1 to 4, execute step 5 for iteration, and display a complete domain technology tree about quantum chips in the form of a multi-branch tree.

[0098] Meanwhile, the nodes in the technology tree include downstream technology nodes and dependent technology nodes, clearly showing the technological development path. For example, downstream technology nodes for quantum chips include: photonic chip methods, quantum dots, semiconductor qubits, metamaterial antennas, and quantum coherent receivers. Dependent technologies for quantum chips include: qubits, superconducting quantum, silicon electronics, and silicon nitride photonics. Furthermore, all nodes generated by this method can be added, deleted, and have their details displayed, such as... Figure 2 As shown, the details include the source of the search content corresponding to this node and an explanation of the domain's overall model.

[0099] This invention provides a domain technology tree generation method based on large-scale model knowledge reasoning. By using this method to mine and generate technology trees for key technologies, the understanding of the key technology context in a domain can be effectively improved. Specifically, the method of this invention is used to retrieve text related to key technologies, and a domain large-scale model is used to judge the high-value content of candidate retrieval content. Furthermore, the domain large-scale model is used to generate triples of downstream technologies and dependent technologies that are closely related to the content, thereby accurately constructing a complete domain technology tree related to key technologies.

[0100] Furthermore, embodiments of the present invention also provide a domain technology tree generation device based on large model knowledge reasoning, such as... Figure 5 As shown, it includes: a data retrieval module 51, a high-value content filtering module 52, a technology triple extraction module 53, and a domain technology tree generation module 54; wherein:

[0101] Data retrieval module 51 is used to retrieve candidate search results from the Internet and / or domain knowledge bases using domain technical keywords;

[0102] The high-value content filtering module 52 is used to construct a first task instruction using the candidate search content and the domain technical keywords, and call the domain big model to judge the relevance between the candidate search content and the technical keywords in the first task instruction to obtain high-value candidate search content.

[0103] The technology triple extraction module 53 is used to construct a second task instruction using the high-value candidate search content and the domain technology keywords, and to call the domain large model to extract domain technology triples from the high-value candidate search content of the second task instruction. The domain technology triples use the domain technology keywords as the subject, any word with "inclusion" semantics within the relation set as the predicate, and downstream technology keywords or dependent technology keywords as the object.

[0104] Domain technology tree generation module 54 is used to generate a domain technology tree using the extracted domain technology triples.

[0105] Preferably, it also includes:

[0106] The domain technology tree display module 55 is used to trigger the data retrieval module when it is determined that the extracted downstream technology keywords need to be added to the domain technology keywords.

[0107] Specifically, the data retrieval module 51 includes: a retrieval unit, a sorting unit, and a filtering unit;

[0108] The retrieval unit is used to retrieve the Internet and / or domain knowledge bases using domain technical keywords to obtain the retrieval content;

[0109] The sorting unit is used to sort the search content from high to low according to the relevance of the search content to the domain technical keywords, and take the first set number of search content as candidate search content.

[0110] The filtering unit is used to perform deduplication filtering and format parsing on candidate search content to generate candidate search content in a uniform format.

[0111] Preferably, the high-value content filtering module 52 includes: a traversal unit, a first prompt question composition unit, a matching unit, and a high-value content construction unit;

[0112] The traversal unit is used to traverse the candidate search content in the candidate search dataset, where all candidate search content constitutes the candidate search dataset.

[0113] The first prompt question unit is used to combine candidate search content with their corresponding domain technical keywords to form key technology pairs, and then use prompting engineering to form the key technology pairs into the first prompt question.

[0114] The matching unit is used to take the first prompt question as the first task instruction, input it into the domain big model, and get the answer about whether the candidate search content matches the domain technical keywords;

[0115] The high-value content building unit is used to select candidate search content whose answers match as high-value candidate search content.

[0116] Specifically, the technology triple extraction module 53 includes: a relation set construction unit, a traversal unit, a second prompt question composition unit, and a technology triple generation unit;

[0117] The relation set construction unit is used to construct a relation set, wherein the relation set contains predicate keywords representing hierarchical relationships;

[0118] The traversal unit is used to traverse the high-value candidate retrieval content in the high-value candidate dataset;

[0119] The second prompt question component unit is used to form a second prompt question based on high-value candidate search content, using prompt engineering, with the domain technical keywords corresponding to the high-value candidate search content as the subject and any word with the semantic meaning of "inclusion" in the relation set as the predicate;

[0120] The technical triple generation unit is used to input the second prompt question as a second task instruction into the domain big model to determine whether there is a predicate in the high-value candidate content that conforms to the technical relationship of the triple. When the determination result is that there is, a matching domain technical triple is generated as the answer.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A domain technology tree generation method based on large model knowledge reasoning, comprising: Step 1: Use domain-specific technical keywords to search the Internet and / or domain-specific knowledge bases to obtain candidate search results; Step 2: Construct a first task instruction using the candidate search content and the domain technical keywords, and call the domain big model to judge the relevance between the candidate search content and the technical keywords in the first task instruction to obtain high-value candidate search content; Step 3: Construct a second task instruction using the high-value candidate search content and the domain technology keywords, and call the domain big model to extract domain technology triples from the high-value candidate search content of the second task instruction. The domain technology triples take the domain technology keywords as the subject, any word with "inclusion" semantics in the relation set as the predicate, and downstream technology keywords or dependent technology keywords as the object. Step 4: Generate a domain technology tree using the extracted domain technology triples; Specifically, step 3 includes: Step 3-1: Construct a set of relations, which contains predicate keywords representing hypernym-hypernym relationships; Step 3-2: Traverse the high-value candidate search content in the high-value candidate dataset, where all high-value candidate search content constitute the high-value candidate dataset; Step 3-3: Based on high-value candidate search content, using suggestion engineering, the second suggestion question is formed by using the domain technical keywords corresponding to the high-value candidate search content as the subject and any word with the semantic meaning of "inclusion" in the relation set as the predicate. Steps 3-4: Use the second prompt question as the second task instruction, input the domain big model to determine whether there are predicates in the high-value candidate content that conform to the triplet technical relationship, and generate the matching domain technical triplet as the answer when the judgment result is that it exists.

2. The domain technology tree generation method based on large model knowledge reasoning according to claim 1, characterized in that, Also includes: Step 5: When it is determined that the downstream technical keywords extracted this time need to add child nodes, the downstream technical keywords extracted this time are used as domain technical keywords, and the process jumps to step 1.

3. The domain technology tree generation method based on large model knowledge reasoning according to claim 1, characterized in that, Step 1 includes: Step 1-1: Use domain-specific technical keywords to search the Internet and / or domain-specific knowledge bases to obtain the search results; Step 1-2: Sort the search results from highest to lowest according to their relevance to the domain's technical keywords, and select a set number of the top-ranked search results as candidate search results; Steps 1-3: Perform deduplication filtering and format parsing on the candidate search content to generate candidate search content in a uniform format.

4. The domain technology tree generation method based on large model knowledge reasoning according to any one of claims 1-3, characterized in that, Step 2 includes: Step 2-1: Traverse the candidate search content in the candidate search dataset, where all candidate search content constitutes the candidate search dataset; Step 2-2: Combine candidate search results with their corresponding domain technical keywords to form key technology pairs, and use prompting engineering to form the first prompt question from the key technology pairs; Steps 2-3: Input the first prompt question as the first task instruction into the domain big model to get the answer regarding whether the candidate search content matches the domain technical keywords; Steps 2-4: Select candidate search terms that match the answer as high-value candidate search terms.

5. A domain technology tree generation device based on large-model knowledge reasoning, characterized in that, include: The data retrieval module is used to search the Internet and / or domain knowledge bases using domain technical keywords to obtain candidate search results; The high-value content filtering module is used to construct a first task instruction using the candidate search content and the domain technical keywords, and call the domain big model to judge the relevance between the candidate search content and the technical keywords in the first task instruction to obtain high-value candidate search content. The technology triple extraction module is used to construct a second task instruction using the high-value candidate search content and the domain technology keywords, and to call the domain large model to extract domain technology triples from the high-value candidate search content of the second task instruction. Each domain technology triple has a domain technology keyword as the subject, any word with "inclusion" semantics within the relation set as the predicate, and downstream technology keywords or dependent technology keywords as the object. The domain technology tree generation module is used to generate a domain technology tree using the extracted domain technology triples; The technology triple extraction module includes: a relation set construction unit, a traversal unit, a second prompt question composition unit, and a technology triple generation unit; The relation set construction unit is used to construct a relation set, wherein the relation set contains predicate keywords representing hierarchical relationships; The traversal unit is used to traverse the high-value candidate retrieval content in the high-value candidate dataset, wherein all high-value candidate retrieval content constitutes the high-value candidate dataset. The second prompt question unit is used to form a second prompt question based on high-value candidate search content, using prompt engineering, with the domain technical keywords corresponding to the high-value candidate search content as the subject and any word with the semantic meaning of "inclusion" in the relation set as the predicate; The technical triple generation unit is used to input the second prompt question as a second task instruction into the domain big model to determine whether there is a predicate in the high-value candidate content that conforms to the technical relationship of the triple. When the determination result is that there is, a matching domain technical triple is generated as the answer.

6. The domain technology tree generation device based on large model knowledge reasoning according to claim 5, characterized in that, Also includes: The domain technology tree display module is used to trigger the data retrieval module when it is determined that the extracted downstream technology keywords need to be added to the domain technology keywords.

7. The domain technology tree generation device based on large model knowledge reasoning according to claim 5, characterized in that, The data retrieval module includes: a retrieval unit, a sorting unit, and a filtering unit; The retrieval unit is used to retrieve the Internet and / or domain knowledge bases using domain technical keywords to obtain the retrieval content; The sorting unit is used to sort the search content from high to low according to the relevance of the search content to the domain technical keywords, and take the first set number of search content as candidate search content. The filtering unit is used to perform deduplication filtering and format parsing on candidate search content to generate candidate search content in a uniform format.

8. The domain technology tree generation device based on large model knowledge reasoning according to any one of claims 5-7, characterized in that, The high-value content filtering module includes: a traversal unit, a first prompt question composition unit, a matching unit, and a high-value content construction unit; The traversal unit is used to traverse the candidate search content in the candidate search dataset, wherein all candidate search content constitutes the candidate search dataset. The first prompt question unit is used to combine candidate search content with its corresponding domain technical keywords to form key technology pairs, and use prompt engineering to form the key technology pairs into the first prompt question; The matching unit is used to input the first prompt question as the first task instruction into the domain big model and obtain an answer regarding whether the candidate search content matches the domain technical keywords. The high-value content construction unit is used to select candidate search content whose answers match as high-value candidate search content.

Images