Method, device and system for generating patent summary

Abstract

The embodiment of the invention provides a patent summary generation method, device and system, and the method comprises the steps: firstly obtaining a target knowledge graph corresponding to a target patent file based on the patent summary generation method; then, the target knowledge graph is used as assistance, the patent summary model is effectively questioned according to the patent content of the target patent file, and a patent summary of the target patent file is obtained; the patent summary model may include a large language model. Therefore, the accurate patent summary can be efficiently and automatically obtained at a low processing cost.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to methods, devices and systems for generating patent outlines. Background Technology

[0002] With the development of large language models, some users have begun to try using large language models to interpret patent documents and extract the corresponding patent summaries.

[0003] Based on an existing method, users input a question containing many technical terms into a large language model at once to ask a summary of the patent document.

[0004] Based on another existing method, users would manually engage in multiple rounds of question-and-answer dialogues with a large language model until they finally elicited the underlying technology from the patent documents, thus obtaining the required patent summary.

[0005] It should be noted that the above introduction to the technical background is only for the purpose of providing a clear and complete explanation of the technical solutions of this application and facilitating understanding by those skilled in the art. It should not be assumed that these technical solutions are known to those skilled in the art simply because they have been described in the background section of this application. Summary of the Invention

[0006] The inventors discovered that, with the first existing method, because users simultaneously input questions containing many technical terms into the large language model, the model often cannot accurately understand the corresponding multiple technical terms at the same time. This leads to discrepancies between the output patent summary and the actual patent document. For example, the resulting patent summary may omit some core technical content.

[0007] Regarding the second existing method, firstly, the multi-turn question-and-answer dialogue process with the large language model relies entirely on the user's subjective factors, placing high demands on the user's personal experience and knowledge reserves. Limited by personal experience and knowledge reserves, users often fail to accurately grasp the best prompts for their questions, thus affecting the questioning effect and resulting in poor accuracy, instability, and a tendency to errors in the final patent summary.

[0008] Meanwhile, the second existing method, in multi-turn question-and-answer dialogues with a large language model, requires users to focus on analyzing and reflecting on the current response text in each round to determine the prompts for the next round of questions. However, most users often find it difficult to maintain focus on analyzing and reflecting on the response text for extended periods, making it impossible to continuously ask in-depth and effective questions. This results in a lengthy and inefficient multi-turn question-and-answer dialogue process.

[0009] Furthermore, both the first and second existing methods require users to manually engage in question-and-answer dialogues with the large language model. The question-and-answer dialogue process with the large language model is itself tedious and complex, increasing the user's workload and the overall processing cost.

[0010] To address one or more of the aforementioned problems, embodiments of this application provide a method, apparatus, and system for generating patent summaries, which can efficiently and automatically obtain patent summaries of patent documents with high accuracy and minimal error at a lower processing cost.

[0011] According to a first aspect of the embodiments of this application, a method for generating a patent summary is provided, the method comprising: obtaining a target knowledge graph corresponding to a target patent document; and, based on the target knowledge graph, querying a patent summary model for the patent content of the target patent document to obtain a patent summary of the target patent document; wherein the patent summary model includes a large language model.

[0012] According to a second aspect of the present application, a patent summary generation apparatus is provided, the patent summary generation apparatus comprising: a memory storing a computer program; and a processor that executes the computer program to implement any of the patent summary generation methods described in the first aspect of the present application.

[0013] According to a third aspect of the embodiments of this application, a patent summary generation system is provided, the patent summary generation system including the patent summary generation device described in the second aspect of the embodiments of this application.

[0014] One of the beneficial effects of this application's embodiments is that by introducing and using a knowledge graph corresponding to the patent document as an aid during the question-and-answer dialogue with the patent summary model, appropriate prompt words can be determined efficiently and accurately, and questions can be automatically asked to the model based on the prompt words to obtain the corresponding patent summary.

[0015] This effectively reduces the user's workload, lowers the overall processing cost, and efficiently and intelligently obtains accurate patent summaries automatically.

[0016] Furthermore, the target knowledge graph includes at least the technical elements in the target patent document and the relationships between the technical elements.

[0017] Therefore, a knowledge graph with rich and comprehensive data information about patent documents can be constructed, which is suitable for asking questions about the patent summary.

[0018] Furthermore, the step of "obtaining the target knowledge graph corresponding to the target patent document" includes: generating a target knowledge graph about the target patent document based on the patent text features of the target patent document; wherein, the patent text features include: patent text structural features and patent text vocabulary features.

[0019] Therefore, by effectively utilizing patent text features such as structural features and vocabulary features, a knowledge graph with good performance can be constructed efficiently and accurately for the patent document. Subsequently, questions can be asked more accurately on the patent summary model based on the knowledge graph, resulting in a patent summary with relatively higher accuracy and smaller error.

[0020] Furthermore, the step of "obtaining the target knowledge graph corresponding to the target patent document" includes: performing entity recognition on the target patent document to determine the entity objects in the target patent document; wherein, the entity objects include at least technical elements; extracting relationships from the target patent document to obtain the association relationships between the entity objects; and constructing the target knowledge graph corresponding to the target patent document based on the entity objects and the association relationships between the entity objects.

[0021] Therefore, by sequentially extracting and utilizing entity objects that include at least technical elements from the patent documents, as well as the relationships between these entity objects, a relatively accurate knowledge graph about the patent documents can be efficiently constructed.

[0022] Furthermore, the technical elements include: technical means elements, technical problem elements, and technical effect elements.

[0023] Therefore, by acquiring and utilizing the technical elements that users are more concerned about in the patent summary as entity objects, a knowledge graph with good determination effect for the patent summary can be constructed.

[0024] Furthermore, the step of "generating a target knowledge graph of the target patent document based on the patent text features" includes: dividing the target patent document into multiple modules based on the patent text features; wherein the multiple modules include a background technology module, a specification content module, and an embodiment module; performing entity recognition and relationship extraction on each of the multiple modules to obtain the entity objects of each of the multiple modules and the association relationships between the entity objects; and constructing corresponding technical problem knowledge graphs, technical means knowledge graphs, and technical effect knowledge graphs based on the entity objects of each of the multiple modules and the association relationships between the entity objects.

[0025] Therefore, by fully utilizing the patent text features of patent documents to process patent documents, the required entity objects and the relationships between entity objects can be extracted efficiently and accurately to construct the corresponding knowledge graph, which effectively improves the processing efficiency when constructing the knowledge graph.

[0026] Furthermore, the step "based on the target knowledge graph, ask questions to the patent summary model regarding the patent content of the target patent document to obtain the patent summary of the target patent document" includes: selecting target technical elements that meet the requirements based on the target knowledge graph; generating corresponding prompts based on the selected target technical elements; using the prompts to ask questions to the patent summary model regarding the patent content of the target patent document to obtain corresponding response text; and generating the patent summary of the target patent document based on the response text.

[0027] Therefore, by utilizing knowledge graphs, appropriate prompts can be automatically determined based on the target technical elements that meet the requirements. Then, based on these prompts, more effective questions can be asked about the patent summary model of the patent document, thereby efficiently and accurately determining the patent summary of the patent document.

[0028] Furthermore, the step of "selecting target technical elements that meet the requirements based on the target knowledge graph" includes: selecting at least one target technical element from different knowledge graphs in the target knowledge graph as a target technical element that meets the requirements; wherein, the target knowledge graph includes: a technical problem knowledge graph, a technical means knowledge graph, and a technical effect knowledge graph.

[0029] Therefore, different knowledge graphs can be used in a targeted manner to efficiently and accurately identify the required target technical elements.

[0030] Furthermore, the step of "selecting target technical elements that meet the requirements based on the target knowledge graph" includes: obtaining summary type parameters; selecting target technical elements that meet the requirements based on the summary type parameters and the target knowledge graph; wherein, the summary type parameters include one or more of the following: easy to understand, core summary, comprehensive summary, higher-level summary, and personalized customization.

[0031] Therefore, by introducing and filtering matching target technical elements based on the summary type parameters, the corresponding type of patent summary can be accurately determined to meet the diverse needs of users for patent summaries and improve the user experience.

[0032] Furthermore, the types of prompt words include at least one of the following: text prompts, image prompts, and sound prompts.

[0033] Therefore, it can support asking questions about the patent summary model using various types of prompts, and can more efficiently and conveniently determine the patent summary that meets the requirements through questioning.

[0034] Furthermore, the step "generating a patent summary of the target patent document based on the response text" includes: generating a patent summary of the target patent document using the response text based on the summary template.

[0035] Therefore, it is possible to introduce and utilize summary templates to generate patent summaries that meet the user's customized requirements, thus enabling the user to have a better user experience.

[0036] Furthermore, the summary template includes at least one of the following: the text content of the patent summary arranged in the order of technical problem, technical means and technical effect; the text length of the generated patent summary text content; the language units of the generated patent summary text content; and the language style of the generated patent summary text content.

[0037] Therefore, patent summaries that meet diverse user customization requirements can be generated through summary templates, enabling users to obtain a relatively better user experience.

[0038] Furthermore, the step "using the prompt words to ask questions to the patent summary model regarding the patent content of the target patent document to obtain the patent summary of the target patent document" includes: using the prompt words to ask multiple rounds of questions to the patent summary model regarding the patent content of the target patent document to obtain the patent summary of the target patent document.

[0039] Therefore, by asking multiple questions to the patent summary model, it is possible to conduct progressive and in-depth questioning of the patent documents, resulting in a more accurate patent summary.

[0040] Furthermore, the step of "using the prompt words to conduct multiple rounds of questioning on the patent summary model in response to the patent content of the target patent document" includes: each round of questioning on the patent summary model in response to the patent content of the target patent document includes: obtaining the response text of the previous round; determining the prompt words for the current round based on the response text of the previous round and the target technical elements; using the prompt words for the current round to conduct the current round of questioning on the patent summary model in response to the patent content of the target patent document, and obtaining the response text for the current round.

[0041] Therefore, by effectively utilizing the response text from the previous round, more suitable prompts can be identified for the current round. These prompts can then be used to ask questions more accurately and in-depth, thereby obtaining the required patent summary more efficiently, reducing the number of rounds of question-and-answer dialogue and the time spent manually writing prompts, and shortening the overall processing time.

[0042] Furthermore, the step "based on the response text and target technical elements of the previous round" includes: extracting the main words from the response text of the previous round; querying the target knowledge graph based on the main words and the target technical elements of the previous round to determine the technical elements that meet the similarity requirements; and generating the prompt words for the current round based on the technical elements that meet the similarity requirements.

[0043] Therefore, by utilizing knowledge graphs and based on the dimension of semantic similarity, we can accurately identify the target technical elements that meet the requirements, so as to generate prompt words with better results.

[0044] Furthermore, if the technical elements that meet the similarity requirements include multiple technical elements that meet the similarity requirements, the technical element with the highest similarity can be determined from the multiple technical elements that meet the similarity requirements; and the prompt word for the current round can be generated based on the technical element.

[0045] Therefore, by utilizing knowledge graphs and based on the dimension of semantic similarity, the most suitable technical element can be selected from multiple technical elements as the target technical element to generate the most appropriate prompt words for the current round of questions.

[0046] Furthermore, the step of "determining the prompt words for the current round based on the response text and target technical elements of the previous round" includes: extracting the main words from the response text of the previous round; querying the target knowledge graph based on the main words and the target technical elements of the previous round to determine the related technical elements respectively; determining the technical element with the largest weight among the related technical elements based on the weight of the technical elements; and generating the prompt words for the current round based on the technical elements.

[0047] Therefore, by utilizing knowledge graphs and based on the dimension of technological relevance, we can accurately identify the target technological elements that meet the requirements, thereby generating prompts with better results.

[0048] Furthermore, the step of "determining the prompt words for the current round based on the response text of the previous round and the target technical elements" includes: querying the target knowledge graph based on the response text of the current round to determine whether there is a related next technical element; and ending the multi-round questioning if it is determined that there is no related next technical element.

[0049] Therefore, based on the knowledge graph and the relationships between technical elements, it is possible to accurately determine the timing of the end of multiple rounds of questioning, thus avoiding wasting processing time and resources on ineffective questioning.

[0050] Furthermore, after using the prompts of the current round to ask questions to the patent summary model based on the patent content of the target patent document and obtaining the response text of the current round, semantic matching can be performed using relevant embodiments in the target patent document to detect whether the response text of the current round is a response text that meets the requirements; if it is determined that the response text of the current round is a response text that meets the requirements, the multi-round questioning ends.

[0051] Therefore, based on the relevant embodiments in the patent documents and semantic matching, the timing of the end of multiple rounds of questioning can be accurately determined, avoiding wasting processing time and resources on invalid questions.

[0052] Furthermore, after completing multiple rounds of questioning, the technical problem text, technical means text, and technical effect text can be determined based on the response texts of the multiple rounds of questioning. The technical problem text, technical means text, and technical effect text can be combined to obtain the patent summary of the target patent document.

[0053] Therefore, by combining the response texts of multiple rounds of questions, a more complete, comprehensive, and accurate patent summary can be obtained.

[0054] Furthermore, an initial model based on a large language model can be constructed first; then, using the first patent sample data, the initial model can be trained to obtain a first intermediate model that meets the requirements; next, using the second patent sample data, the first intermediate model can be trained to obtain a second intermediate model that meets the requirements; wherein, the second patent sample data is labeled with corresponding patent summaries; finally, by fine-tuning the second intermediate model and adding a preset loop guidance mechanism, a patent summary model that meets the requirements can be obtained.

[0055] Therefore, based on the large language model, different types of patent sample data can be used for training in different ways; then the model can be fine-tuned and a preset loop guidance mechanism can be added to effectively shorten the training time and reduce the training cost of the model, resulting in a patent summary model that is suitable for generating patent summaries for patent documents, has good application effects, and supports progressive multi-round questioning.

[0056] Specific embodiments of this application are disclosed in detail with reference to the following description and accompanying drawings, indicating how the principles of this application can be adopted. It should be understood that the embodiments of this application are not limited in scope. Within the spirit and scope of the appended claims, embodiments of this application include many changes, modifications, and equivalents.

[0057] The feature information described and illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, combined with feature information in other embodiments, or substituted for feature information in other embodiments.

[0058] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, whole, step, or component, but does not exclude the presence or addition of one or more other features, wholes, steps, or components. Attached Figure Description

[0059] The above and other objects, features and advantages of the embodiments of this application will become more apparent from the following detailed description taken in conjunction with the accompanying drawings, in which:

[0060] Figure 1 This is a flowchart illustrating a method for generating a patent summary according to an embodiment of this application.

[0061] Figure 2 This is a flowchart illustrating the process of generating a target knowledge graph using the patent summary generation method of an embodiment of this application in a scenario example.

[0062] Figure 3 This is another flowchart illustrating the generation of a target knowledge graph using the patent summary generation method of an embodiment of this application in a scenario example.

[0063] Figure 4 This is a schematic diagram of a process for generating a patent summary using the patent summary generation method of an embodiment of this application in a scenario example;

[0064] Figure 5 This is a flowchart illustrating the process of generating a patent summary using an embodiment of this application in a scenario example, for the current round of questioning.

[0065] Figure 6 This is a flowchart illustrating the process of training a patent summary model using the patent summary generation method of this application in a scenario example.

[0066] Figure 7 This is another flowchart illustrating the process of training a patent summary model using the patent summary generation method of this application embodiment in a scenario example;

[0067] Figure 8 This is a schematic diagram of a process flow illustrating the patent summary generation method of this application embodiment applied in a scenario example, utilizing a core generalization-based guiding mechanism of the patent summary model.

[0068] Figure 9This is a detailed processing flowchart illustrating the patent summary generation method of this application embodiment applied in a scenario example, utilizing a core summary-based guiding mechanism of the patent summary model.

[0069] Figure 10 This is a detailed processing flowchart illustrating the patent summary generation method of this application embodiment applied in a scenario example, utilizing a core summary-based guiding mechanism of the patent summary model.

[0070] Figure 11 This is a schematic diagram of a process flow for generating a patent summary using an easy-to-understand guidance mechanism based on a patent summary model, applied in a scenario example of an embodiment of this application.

[0071] Figure 12 This is a schematic diagram of the structure of the device for generating a patent summary according to an embodiment of this application;

[0072] Figure 13 This is a schematic diagram of a patent summary generation apparatus according to an embodiment of this application. Detailed Implementation

[0073] Referring to the accompanying drawings, the foregoing and other features of this application will become apparent from the following description. Specific embodiments of this application are specifically disclosed in the description and drawings, illustrating partial embodiments in which the principles of this application can be employed. It should be understood that this application is not limited to the described embodiments, but includes all modifications and equivalents falling within the scope of the appended claims.

[0074] The method, apparatus, and system for generating patent summaries of embodiments of this application will now be described in conjunction with the accompanying drawings.

[0075] First aspect of the embodiments

[0076] An embodiment of the first aspect of this application provides a method for generating a patent summary. Figure 1 This is a flowchart illustrating a method for generating a patent summary according to an embodiment of this application. Figure 1 As shown, the method for generating this patent summary includes:

[0077] S101: Obtain the target knowledge graph corresponding to the target patent document;

[0078] S102: Based on the target knowledge graph, ask questions to the patent summary model regarding the patent content of the target patent document to obtain the patent summary of the target patent document; wherein, the patent summary model includes a large language model.

[0079] Based on the above embodiments, before implementation, a patent summary model suitable for processing patent documents can be trained using a large language model. During implementation, a target knowledge graph corresponding to the target patent document can be obtained first; then, using the target knowledge graph as an aid, valid questions about the target patent document can be posed to the patent summary model to obtain the corresponding patent summary. This approach allows for the efficient and automatic acquisition of patent summaries of target patent documents with high accuracy and low error; it also effectively reduces the user's workload and lowers the overall processing cost.

[0080] In this embodiment, the target patent document can specifically be a patent document from which a patent summary is to be extracted. In other embodiments, the solution provided in this application can also be used to generate summaries for other types of target documents. For example, the solution provided in this application can also be used to process technical documents such as papers, journal articles, and lecture manuscripts to generate corresponding document summaries.

[0081] In this application embodiment, the patent document can be of various types, such as an invention patent, a utility model patent, or a PCT patent.

[0082] In this embodiment of the application, the patent document can be a patent document published in various forms, such as patent documents published or announced by the State Intellectual Property Office or other official organizations, such as the published text of a patent application, the announced text of a patent grant, or patent documents that can be retrieved from various databases. In this embodiment of the application, the patent document can be a patent document written in Chinese, such as an invention patent, utility model patent, or PCT patent applied for, published, or announced in China. However, the patent document can also be a patent document written in other languages ​​(such as English or Japanese), and this embodiment of the application does not limit the language of the patent document.

[0083] In some embodiments, the patent documents may include the following parts: background art, invention description or utility model description (hereinafter collectively referred to as specification content), and textual content such as embodiments (e.g., detailed implementation methods). Furthermore, the patent documents may also include other relevant content such as abstracts and specification drawings.

[0084] Specifically, the aforementioned target patent documents can be patent documents based on different text formats. For example, the target patent documents can be patent documents in Word format, PDF format, Excel format, or the patent content itself, etc.

[0085] Specifically, the aforementioned target patent documents can be patent documents involving different technical fields. For example, the aforementioned target patent documents can be patent documents involving the field of air conditioning control, or patent documents involving the field of algorithms, or patent documents involving the field of intelligent manufacturing, and so on.

[0086] It should be noted that the target patent documents listed above are merely illustrative. In specific implementations, depending on the specific application scenario, the aforementioned target patent documents may also include patent documents based on other text formats or involving other technical fields. This application does not limit this aspect.

[0087] The aforementioned patent summary can be understood as a summary text of the patent content of the invention or utility model claimed in the patent document, such as a summary that embodies the core content of the patent document.

[0088] For example, the patent summary could be a text that outlines the technical problem the patent document aims to solve, the technical means employed to address that problem, and the technical effects achieved based on those means.

[0089] In some embodiments, the aforementioned patent summary model can be understood as an algorithmic model pre-trained based on a large language model, adapted for processing patent documents, and capable of automatically determining the patent summary of a patent document through questioning. The training method for the patent summary model will be explained in detail later.

[0090] In some embodiments, the aforementioned target knowledge graph can be specifically understood as a graph data corresponding to a target patent document, consisting of multiple nodes and edges connecting the nodes.

[0091] In some embodiments, the aforementioned target knowledge graph may include at least the technical elements in the target patent document and the relationships between the technical elements.

[0092] Specifically, in the target knowledge graph, a node represents a technical element; the edge between two nodes can represent the relationship between the two corresponding technical elements.

[0093] Specifically, the aforementioned relationships can be causal, hierarchical, inclusive, structural, or otherwise. Of course, the relationships listed above are merely illustrative. In practice, depending on the specific scenario and processing requirements, the aforementioned relationships may include other types of relationships. This application's embodiments do not limit this.

[0094] In addition, the aforementioned target knowledge graph may also include attribute information about the technical element marked on the node (e.g., the name, type, frequency of occurrence, etc. of the technical element corresponding to the node), and / or attribute information about the corresponding association relationship marked on the edge (e.g., the name, type, degree of association, etc. of the association relationship).

[0095] Based on the above embodiments, a target knowledge graph can be obtained for the target patent document, which contains both the technical elements in the target patent document and the relationships between the technical elements. This data is relatively rich and suitable for subsequent questioning of the patent summary model.

[0096] In some embodiments, the above-mentioned acquisition of the target knowledge graph corresponding to the target patent document may specifically include: generating a target knowledge graph about the target patent document based on the patent text features of the target patent document; wherein, the patent text features include: patent text structural features and patent text vocabulary features.

[0097] Specifically, since patent documents are legal documents, they must be drafted strictly according to fixed patent document drafting rules and meet corresponding normative requirements. Therefore, from a macro perspective, patent documents usually have relatively obvious patent text structure characteristics: they contain multiple fixed text contents such as background technology content, specification content (e.g., invention content, utility model content), and embodiment content (e.g., specific implementation methods); and the text structure between the above multiple text contents is relatively clear, and different text contents mostly contain corresponding identifying words, etc.

[0098] For example, patent documents typically contain different textual content in sequence, such as background art content, specification content, and embodiment content. Furthermore, a separate line often appears between the background art content and the specification content, labeled "Summary of the Invention," and between the specification content and the embodiment content, labeled "Detailed Description of Embodiments." Additionally, within the embodiment content, different embodiments are separated by the labels "In one embodiment" or "In some embodiments."

[0099] Furthermore, according to the relevant patent document drafting rules and specifications, the background technology content will focus more on the technical problem to be solved; while the specification content and the embodiment content will focus more on the technical means adopted and the technical effects achieved.

[0100] At the same time, patent documents usually have relatively obvious characteristics of patent text vocabulary at the micro level: the text vocabulary used to describe the object (including: the executing entity, the subject of execution, steps and actions, etc.) must be clear and unambiguous; and for the same object, there may often be corresponding higher-level, more general text vocabulary and relatively lower-level, more specific text vocabulary at the same time; as well as relatively common keywords in patent documents (e.g., "this invention", "this utility model", etc.).

[0101] Based on the above, in this embodiment, by acquiring and utilizing the patent features of the target patent document, the target patent document can be processed more specifically, thereby obtaining a target knowledge graph with higher accuracy and better usage effect relatively efficiently.

[0102] In some embodiments, Figure 2 This is a flowchart illustrating the process of generating a target knowledge graph using the patent summary generation method of this application in a scenario example. Figure 2 As shown, the acquisition of the target knowledge graph corresponding to the target patent document, in specific implementation, may include the following:

[0103] S201: Perform entity recognition on the target patent document to determine the entity objects in the target patent document; wherein, the entity objects include at least technical elements;

[0104] S202: By extracting relationships from the target patent documents, the association relationships between the entity objects are obtained;

[0105] S203: Based on the entity objects and the relationships between them, construct a target knowledge graph corresponding to the target patent document.

[0106] The aforementioned entity objects can be understood as independent objects existing within the target patent document, serving as the basic units constituting the knowledge graph. Entity objects can be data objects that can be used to generate a patent summary of the target patent document.

[0107] Specifically, the aforementioned entities include at least technical elements.

[0108] In addition, the aforementioned entity objects may also include key indicators used to indicate the invention or utility model claimed in the target patent document, such as "method", "equipment", "apparatus", etc.

[0109] Based on the above embodiments, by identifying entity objects in the target patent document, accurately obtaining and using relevant technical elements and the relationships between technical elements, a target knowledge graph with good effect can be constructed more accurately for the target patent document.

[0110] In some embodiments, the aforementioned technical elements may specifically include: technical means elements, technical problem elements, and technical effect elements, etc.

[0111] Based on the above embodiments, by acquiring and using technical means elements, technical problem elements, and technical effect elements as technical elements, a target knowledge graph suitable for questioning patent summary models can be constructed.

[0112] Specifically, the aforementioned technical elements can be data elements related to the technical means, such as "NLP technology", "OCR recognition", and "clustering algorithm".

[0113] The aforementioned technical problem elements can be data elements related to technical problems, such as "high cost", "poor user experience", "large error" etc.

[0114] The aforementioned technical effect elements can be data elements related to technical problems, such as "reducing costs", "improving user experience", and "increasing accuracy".

[0115] In some embodiments, when implemented, NLP (Natural Language Processing) technology can be used to identify entities that meet the requirements and the relationships between entities by performing entity recognition on the text content in the target patent document.

[0116] Specifically, natural language processing libraries in Python (such as NLTK and SPACY) can be used to perform entity recognition on the target patent document to identify the entity objects in the target patent document.

[0117] In practice, a text semantic recognition model for patent documents can be constructed first based on NLP algorithms. Then, using this text semantic recognition model and keyword templates related to patent documents, text semantic recognition and semantic comparison can be performed on each text paragraph in the target patent document to identify the entity objects in the target patent document and mark them accordingly.

[0118] After identifying the entity objects, the text semantic recognition model described above can be used to extract the relationships between entity objects in the following way: Search the current text paragraph containing the current entity object, as well as neighboring text paragraphs adjacent to the current text paragraph, to identify neighboring entity objects; then use the text semantic recognition model to perform semantic recognition analysis on the current text paragraph and the text paragraphs containing at least one entity object from both the current entity object and neighboring entity objects, to determine whether a relationship exists between the current entity object and the neighboring entity objects. If a relationship is determined, the specific type of relationship between the two entities can be determined based on the semantic recognition analysis results.

[0119] Based on the above embodiments, by introducing and utilizing NLP algorithms, it is possible to automatically determine the entity objects that meet the requirements and the relationships between the entity objects in a relatively efficient and precise manner.

[0120] In some embodiments, after identifying the entity objects in the target patent document and the relationships between them, nodes corresponding to the entity objects can be drawn first using a graph database (e.g., NEO4J) or graph algorithm; then, the nodes corresponding to the entity objects with relationships can be connected by connecting edges to obtain the target knowledge graph.

[0121] After the target knowledge graph is constructed, it can be visualized using graphical interface tools or programming libraries for users to browse and use.

[0122] In some embodiments, after the target knowledge graph is constructed, it can be displayed to the user for verification. When the user finds an error in the target knowledge graph, they can initiate a modification command for it; this modification command can carry user-defined modification parameters. Accordingly, the user can respond to the modification command and modify the target knowledge graph using the aforementioned modification parameters to obtain a target knowledge graph with higher accuracy and lower error.

[0123] In some embodiments, after constructing the target knowledge graph in the manner described above, the attribute information of the entity objects and the attribute information of the association relationships can be determined based on the target patent documents; then, based on the attribute information of the entity objects and the attribute information of the association relationships, the corresponding nodes and edges in the target knowledge graph are labeled to obtain a target knowledge graph with relatively richer and more comprehensive data information.

[0124] In some embodiments, before performing entity identification on the target patent document, the target patent document may be preprocessed, such as by data cleaning, to ensure the accuracy and consistency of the target patent document used subsequently.

[0125] Specifically, the entire target patent document can be traversed to detect whether there is duplicate text content. After detecting duplicate text content, the duplicate text content in the target patent document is removed to avoid wasting subsequent processing time and resources on repeatedly processing the duplicate text content.

[0126] Specifically, it can also detect whether there are obvious text errors in the target patent document (e.g., typos, missing words, or missing punctuation marks) according to preset text rules; after detecting the above obvious text errors, it can automatically correct the text errors in the target patent document according to preset text rules to avoid the above text errors from affecting subsequent processing.

[0127] Specifically, it can also check whether the data used in the target patent document (e.g., English names, tables, formulas, etc.) conforms to the corresponding format requirements according to the patent document writing rules. After detecting data that does not conform to the format requirements, it can automatically adjust the data format according to the patent document writing rules so that the adjusted data conforms to the corresponding format requirements, so as to avoid inconsistencies in data format that may affect the subsequent identification and processing of the data.

[0128] Of course, it should be noted that the preprocessing methods listed above are merely illustrative. In actual implementation, other types of preprocessing may be introduced depending on the specific circumstances and processing requirements. This application does not limit the scope of such preprocessing methods.

[0129] Based on the above embodiments, by preprocessing the target patent document, errors in the target patent document can be detected and eliminated in advance, ensuring the accuracy and consistency of the target patent document and avoiding interference with subsequent processing.

[0130] In some embodiments, Figure 3 This is another flowchart illustrating the generation of a target knowledge graph using the patent summary generation method of this application's embodiments in a scenario example. For example... Figure 3 As shown, the acquisition of the target knowledge graph corresponding to the target patent document, in specific implementation, may include the following:

[0131] S301: Based on the patent text features, the target patent document is divided into multiple modules; wherein, the multiple modules include a background technology module, a specification content module, and an embodiment module;

[0132] S302: Perform entity recognition and relationship extraction on the multiple modules respectively to obtain the entity objects of each of the multiple modules and the association relationships between the entity objects;

[0133] S303: Based on the entity objects of each of the multiple modules and the relationships between the entity objects, construct corresponding technical problem knowledge graphs, technical means knowledge graphs, and technical effect knowledge graphs.

[0134] The aforementioned technical problem knowledge graph can be understood as a sub-graph primarily used to reflect the technical problems that need to be solved in the target patent document, as well as related content of the technical problems. The aforementioned technical means knowledge graph can be understood as a sub-graph primarily used to reflect the technical means employed in the target patent document, as well as related content of the technical means. The aforementioned technical effect knowledge graph can be understood as a sub-graph primarily used to reflect the technical effects achieved in the target patent document based on the employed technical means, as well as related content of the technical effects.

[0135] In practice, firstly, based on the aforementioned patent text characteristics, the target patent document can be divided into several different modules, such as the background technology module, the specification content module, and the embodiment module, by searching and utilizing the identifiers in the target patent document.

[0136] Then, by performing entity recognition on the aforementioned background technology module, specification content module, embodiment module, and other different modules, the entity objects in each module are determined.

[0137] Furthermore, relationships are extracted separately for each module to obtain the association relationships between entity objects in each module.

[0138] Next, a technical problem knowledge graph can be constructed based on the entity objects in the background technology module and the relationships between them; a technical means knowledge graph can be constructed by combining the entity objects in the instruction manual content module and the embodiment module and the relationships between them; and a technical effect knowledge graph can be constructed by combining the entity objects in the instruction manual content module and the embodiment module and the relationships between them. The technical problem knowledge graph corresponds to the background technology module in the target patent document; the technical means knowledge graph corresponds to the combination of the instruction manual content module and the embodiment module in the target patent document; and the technical effect knowledge graph corresponds to the combination of the instruction manual content module and the embodiment module in the target patent document.

[0139] In addition, the above-mentioned technical problem knowledge graph, technical means knowledge graph, and technical effect knowledge graph can be combined to obtain a target knowledge graph for the target patent document.

[0140] Based on the above embodiments, patent text features can be effectively utilized to construct a more refined target knowledge graph suitable for determining the patent summary.

[0141] In some embodiments, Figure 4 This is a flowchart illustrating the process of generating a patent summary using the patent summary generation method of an embodiment of this application in a scenario example. Figure 4 As shown, the above-mentioned method obtains a patent summary of the target patent document by querying the patent summary model based on the target knowledge graph and the patent content of the target patent document. In specific implementation, it may include the following:

[0142] S401: Based on the target knowledge graph, select the target technical elements that meet the requirements;

[0143] S402: Generate corresponding prompt words based on the selected target technical elements;

[0144] S403: Using the prompt words, ask questions to the patent summary model about the patent content of the target patent document, and obtain the corresponding response text;

[0145] S404: Generate a patent summary of the target patent document based on the response text.

[0146] In practice, corresponding questions can be generated based on prompt words; then, the questions can be used to ask the patent summary model questions about the relevant patent content in the target patent document in order to obtain the corresponding response text.

[0147] In practice, the technical problem text, technical means text, and technical effect text can be determined based on the response text; then, the technical problem text, technical means text, and technical effect text can be combined to obtain the patent summary of the target patent document.

[0148] Based on the above embodiments, a target knowledge graph can be used to accurately select suitable target technical elements; then, based on the above target technical elements, prompt words with better effects can be generated and used to effectively ask questions to the patent summary model, so as to efficiently obtain a patent summary with better accuracy and smaller error.

[0149] In some embodiments, the above-mentioned selection of target technical elements that meet the requirements based on the target knowledge graph may specifically include: selecting at least one target technical element from different knowledge graphs in the target knowledge graph as a target technical element that meets the requirements; wherein, the target knowledge graph includes: a technical problem knowledge graph, a technical means knowledge graph, and a technical effect knowledge graph.

[0150] This takes into account that different knowledge graphs in the target knowledge graph are generated based on different modules in the target patent document, which leads to differences in the main content represented by different knowledge graphs.

[0151] Specifically, a technical problem knowledge graph focuses more on representing technical problems and related content. A technical means knowledge graph focuses more on representing technical means and related content. A technical effect knowledge graph focuses more on representing technical effects and related content.

[0152] Therefore, in this embodiment, different knowledge graphs can be distinguished, and different types of technical elements can be selected as target technical elements that meet the requirements for different knowledge graphs.

[0153] Specifically, technical elements related to the technical problem can be selectively selected from the technical problem knowledge graph as target technical elements (which can be referred to as the first type of target technical elements); technical elements related to the technical means can be selectively selected from the technical means knowledge graph as target technical elements (which can be referred to as the second type of target technical elements); and technical elements related to the technical effects can be selectively selected from the technical effect knowledge graph as target technical elements (which can be referred to as the third type of target technical elements).

[0154] Furthermore, by using the first type of target technical elements to generate corresponding prompts and asking questions to the patent summary model, the technical problem text describing the technical problem of the target patent document can be obtained more efficiently and accurately. Similarly, by using the second type of target technical elements to generate corresponding prompts and asking questions to the patent summary model, the technical means text describing the technical methods of the target patent document can be obtained more efficiently and accurately. Finally, by using the third type of target technical elements to generate corresponding prompts and asking questions to the patent summary model, the technical effect text describing the technical effect of the target patent document can be obtained more efficiently and accurately.

[0155] In some embodiments, technical elements related to the technical problem can be selectively selected from the technical problem knowledge graph as first-category target technical elements; then, corresponding prompts can be generated using the first-category target technical elements to ask questions to the patent summary model; then, based on the response text corresponding to the question, or directly based on the first-category target technical elements, technical elements that conform to the technical means of the technical problem can be selected from the technical means knowledge graph as second-category target technical elements; then, corresponding prompts can be generated using the second-category target technical elements to ask questions to the patent summary model; then, based on the response text corresponding to the question, or directly based on the first-category target technical elements, technical elements that conform to the technical effects of the technical problem and technical means can be selected from the technical effects knowledge graph as third-category target technical elements; finally, corresponding prompts can be generated using the third-category target technical elements to ask questions to the patent summary model.

[0156] Based on the above embodiments, by distinguishing different knowledge graphs, suitable target technical elements can be determined more accurately and efficiently.

[0157] In some embodiments, the above-mentioned filtering of target technical elements that meet the requirements based on the target knowledge graph may further include: obtaining summary type parameters; and filtering target technical elements that meet the requirements based on the summary type parameters and the target knowledge graph. The summary type parameters may include one or more of the following: easy to understand, core summary, comprehensive summary, higher-level summary, personalized customization, etc.

[0158] Specifically, in order to meet the diverse summary needs of users, the patent summary generation method provided in this application can also support the generation of various types of patent summaries, such as: easy-to-understand patent summaries, core summary patent summaries, comprehensive summary patent summaries, higher-level summary patent summaries, and personalized customized patent summaries.

[0159] The aforementioned easy-to-understand patent summaries mostly use relatively general expressions (e.g., common terms) instead of relatively abstract expressions (e.g., patent terms) to make the patent summaries relatively easy to understand.

[0160] In the above-mentioned general patent summary, the descriptions of the invention points and / or technical keywords in the target patent documents are mostly replaced by descriptions of the higher level.

[0161] The aforementioned core summary patent outlines, in terms of content, mostly summarize the most core inventive points in the target patent document.

[0162] The aforementioned comprehensive patent summary will generally summarize all the inventive points involved in the target patent document.

[0163] Most of the aforementioned personalized patent summaries are generated based on user-defined rules to meet individual user requirements. For example, user-specific terminology can be used in the patent summary to describe the inventive points and / or technical keywords in the target patent document, according to user-defined rules. Another example is generating a patent summary containing only the two most critical inventive points, based on user requirements.

[0164] Of course, it should be noted that the summary type parameters listed above are only illustrative. In specific implementations, depending on the specific application scenario and processing requirements, the above summary type parameters may also include other type parameters. This application does not limit this aspect.

[0165] Specifically, the aforementioned patent summary model also incorporates several pre-defined cyclical guidance mechanisms. These cyclical guidance mechanisms can be understood as a progressive, multi-round questioning mechanism.

[0166] Specifically, the aforementioned preset cyclical guidance mechanisms may include: easy-to-understand guidance mechanisms, core summary guidance mechanisms, comprehensive summary guidance mechanisms, higher-level summary guidance mechanisms, and personalized guidance mechanisms. Each preset cyclical guidance mechanism corresponds to at least one summary type.

[0167] Accordingly, in specific implementation, a matching preset loop guidance mechanism can be determined from multiple preset loop guidance mechanisms based on the summary type parameter; then, based on the matching preset loop guidance mechanism and the target knowledge graph, the target technical elements that match the summary type can be selected as the target technical elements that meet the requirements.

[0168] Furthermore, based on the aforementioned target technical elements, targeted questions can be posed to the patent summary model to obtain a patent summary of the corresponding summary type.

[0169] Based on the above embodiments, by obtaining and filtering the summary type parameters and using the matching target technical elements to ask questions, a patent summary of the corresponding summary type can be obtained, thus meeting the diverse summary needs of users.

[0170] In some embodiments, the above-mentioned filtering of target technical elements that meet the requirements based on the target knowledge graph may further include: displaying the target knowledge graph to the user; receiving and responding to the user's selection operation on the target knowledge graph; and determining the technical element selected by the user as the target technical element that meets the requirements based on the selection operation.

[0171] Alternatively, the system can first automatically filter out the target technical elements that meet the requirements, and then display these target technical elements and the target knowledge graph together to the user; the node where the target technical element is located is marked in the target knowledge graph. Then, user feedback data is received; based on the user feedback data, the target technical elements are adjusted; and the adjusted target technical elements are used as the qualified target technical elements.

[0172] In some embodiments, during implementation, corresponding prompt words can be automatically generated based on semantic association of the target technical elements. Alternatively, prompt words can be received from user-defined input based on the target knowledge graph as corresponding prompt words.

[0173] In some embodiments, the types of prompt words include at least one of the following: text prompts, image prompts, sound prompts, etc. Of course, the types of prompt words listed above are merely illustrative. In specific implementations, depending on the specific circumstances and processing requirements, the prompt words may also include other types of prompt words.

[0174] In practice, when the prompt includes an image prompt, a pre-trained OCR recognition model can be used to process the image prompt and convert it into a corresponding text prompt.

[0175] When the prompt includes an audio prompt, a pre-trained speech recognition model can be used to process the audio prompt and convert it into a corresponding text prompt.

[0176] Based on the above embodiments, different types of prompt words can be obtained and used to ask questions about the patent summary model more conveniently and efficiently, and obtain the required response text.

[0177] In some embodiments, the above-described generation of a patent summary of a target patent document based on the response text may specifically include the following: generating a patent summary of the target patent document using the response text based on a summary template.

[0178] The aforementioned summary template can be understood as a combination of rules for generating patent summaries.

[0179] Based on the above embodiments, by following the summary template, the response text obtained from the question can be combined to generate a more standardized, uniform, or personalized patent summary.

[0180] In some embodiments, the above-mentioned summary template may include at least one of the following:

[0181] The text content of a patent summary arranged in the order of technical problem, technical means, and technical effect;

[0182] The text length of the generated patent summary text content;

[0183] The language units of the generated patent summary text content;

[0184] The language style, etc., of the generated patent summary text content.

[0185] In practice, users can customize the above-mentioned summary template; and then generate a patent summary that meets the user's requirements based on the user-customized summary template.

[0186] In a patent summary, there can be one or more (two or more) technical problems, technical means, and technical effects. For example, a generated patent summary can contain one technical problem, two technical means, and three technical effects.

[0187] Specifically, users can control the number of characters in the generated patent summary by setting the text length of the text content of the patent summary generated in the summary template.

[0188] For example, a user can set the text length of the patent summary generated from the summary template to 50 characters. Correspondingly, based on the summary template, a response text with a character count of 50 characters or less can be generated as the patent summary of the target patent document.

[0189] Specifically, users can control the language format of the generated patent summary by setting the language units of the text content in the summary template.

[0190] For example, a user can set the language unit of the text content of the patent summary generated in the summary template to a single sentence. Correspondingly, based on the summary template, a single-sentence content text (e.g., a one-line offer) can be generated by combining the response text with the summary text to serve as the patent summary of the target patent document.

[0191] Specifically, users can control the language style of the generated patent summary by setting the language style of the patent summary generated in the summary template.

[0192] For example, users can set the language style of the patent summary generated from the summary template to a comparative style. Correspondingly, based on the summary template, response texts can be combined and a comparative style with existing technologies can be used to generate corresponding content text as the patent summary of the target patent document.

[0193] The aforementioned language style may also include one or more of the following: simple and direct, easy to understand, rigorous and meticulous, etc.

[0194] Based on the above embodiments, users can modify the relevant parameters in the summary template through custom settings to generate customized patent summaries, thereby meeting users' personalized needs and matching different business scenarios.

[0195] Specifically, different companies have their own internal commonly used terms, which differ from industry-standard terms and patent terminology. Therefore, based on a company's internal commonly used term lookup table, relevant terms in the patent summary can be replaced with the company's internal commonly used terms, thus meeting the company's customized patent summary requirements.

[0196] In some embodiments, the above-mentioned use of the prompt words to ask questions about the patent summary model in response to the patent content of the target patent document, thereby obtaining a patent summary of the target patent document, may include:

[0197] Using the prompt words, the patent summary model is asked multiple rounds of questions based on the patent content of the target patent document to obtain the patent summary of the target patent document.

[0198] In practice, a preset cyclical guidance mechanism can be used to ask multiple progressive questions to the patent summary model using prompts, continuously deepening the questions until a satisfactory response text is obtained. The response texts are then combined to obtain a more accurate patent summary.

[0199] Based on the above embodiments, by asking multiple questions to the patent summary model, the target patent document is explored in depth step by step, resulting in a more accurate patent summary.

[0200] In some embodiments, Figure 5 This is a flowchart illustrating the process of generating a patent summary using an embodiment of this application in a scenario example, illustrating the current round of questioning. Figure 5 As shown, the above-mentioned use of the prompt words to conduct multiple rounds of questioning on the patent summary model based on the patent content of the target patent document can specifically include: each round of questioning on the patent summary model based on the patent content of the target patent document includes:

[0201] S501: Retrieve the text of the previous response;

[0202] S502: Based on the response text and target technical elements of the previous round, determine the prompt words for the current round;

[0203] S503: Using the prompts in the current round, ask the patent summary model questions in the current round based on the patent content of the target patent document, and obtain the response text in the current round.

[0204] The aforementioned target technical elements may be the technical elements identified in the previous round.

[0205] Specifically, based on the response text from the previous round, it can be determined whether the previous round of questioning has ended; once it is determined that the previous round of questioning has ended, the response text and the target technical elements from the previous round can be obtained; then, based on the response text and the target technical elements from the previous round, the target technical elements for the current round can be determined; and based on the target technical elements for the current round, the optimal prompt words for the current round can be determined as the prompt words for the current round; and finally, based on the prompt words for the current round, the questions for the current round can be asked.

[0206] It should be noted that the target technical element identified in the current round can be a new technical element or an old technical element that was identified and used in a previous round (e.g., the previous round, or the third round before that).

[0207] Specifically, for example, based on the response text from the previous round, it can be determined that further in-depth questions can be asked about the target technical elements from the previous round. In this case, the target technical elements from the previous round can be used as the target technical elements for the current round; and based on these target technical elements, the optimal prompts for further in-depth questioning can be determined and used as prompts for the current round of questions.

[0208] Based on the above embodiments, the most suitable optimal prompt word for the current round can be accurately determined, and then the current round question can be asked based on the prompt word to obtain a high-value response text for the current round.

[0209] In some embodiments, the above-mentioned method of determining the prompt words for the current round based on the response text of the previous round and the target technical elements may include the following: extracting the main words from the response text of the previous round; querying the target knowledge graph based on the main words and the target technical elements of the previous round to determine the technical elements that meet the similarity requirements; and generating the prompt words for the current round based on the technical elements that meet the similarity requirements.

[0210] Specifically, one can query the target knowledge graph, calculate the similarity between the subject word and the technical elements associated with the target technical elements in the previous round, and determine the technical elements that meet the similarity requirements as the target technical elements in the current round.

[0211] The aforementioned subject words can be subject words representing the executing subject in the previous round's response text, subject words representing the subject being executed, or other objects.

[0212] In practice, based on the main keywords and the target technical elements from the previous round, the target knowledge graph can be queried to identify technical elements whose semantic similarity to the target technical elements is greater than a preset similarity threshold. These are considered technical elements that meet the similarity requirements, i.e., the target technical elements for the current round. Then, based on these similarity-compliant technical elements, prompts for the current round can be generated to ask questions in the current round.

[0213] Based on the above embodiments, suitable prompt words can be determined as prompt words for the current round based on the semantic similarity dimension, according to the target technical elements and the response text of the previous round.

[0214] In some embodiments, where the technical elements meeting the similarity requirement include multiple technical elements meeting the similarity requirement, the method may further include:

[0215] The technical element with the highest similarity is determined from the multiple technical elements that meet the similarity requirements; and a prompt word for the current round is generated based on the technical element.

[0216] Based on the above embodiments, the optimal prompt word can be determined from multiple technical elements that meet the similarity requirements, and the prompt word for the current round can be determined based on the technical element with the highest similarity.

[0217] In some embodiments, the above-mentioned method of determining the prompt words for the current round based on the response text of the previous round and the target technical elements may further include the following: extracting the main words from the response text of the previous round; querying the target knowledge graph based on the main words and the target technical elements of the previous round to determine the related technical elements respectively; determining the technical element with the largest weight among the related technical elements based on the weight of the technical elements; and generating the prompt words for the current round based on the technical elements.

[0218] The weight of the aforementioned technical element can be determined based on the number of edges connected to that technical element in the target knowledge graph.

[0219] The weight percentage of a technical element can be determined by the ratio of the number of edges connected to that technical element to the total number of edges in the target knowledge graph. Generally, a higher weight percentage for a technical element indicates a greater influence and relative importance of that element in the target patent document.

[0220] Specifically, for example, if the number of edges connected to technical element A is 5, and the total number of edges connected in the knowledge graph is 50, then the weight of technical element A can be determined to be 5, and the weight percentage is 5 / 50 = 0.1.

[0221] In practice, based on the main term and the target technical elements of the previous round, the target knowledge graph can be queried to determine the technical elements that are connected to the nodes corresponding to the main term or the target technical elements of the previous round, and these are considered as related technical elements; then, the prompt term for the current round can be generated based on these related technical elements.

[0222] Furthermore, when multiple related technical elements exist, the weight of each related technical element can be determined based on the target knowledge graph; then, based on the weight, the weight percentage of each related technical element can be calculated; and finally, the technical element with the largest weight percentage can be selected from the multiple related technical elements as the target technical element for the current round to generate the prompt word for the current round.

[0223] In some embodiments, after querying the target knowledge graph based on the subject words and the target technical elements of the previous round to determine the associated technical elements, the method may further include the following: statistically analyzing the frequency of occurrence of associated technical elements based on the target patent documents; selecting the technical element with the highest frequency of occurrence from the associated technical elements, and generating the prompt words for the current round based on the technical element.

[0224] Furthermore, after querying the target knowledge graph based on the main keywords and the target technical elements of the previous round to determine the related technical elements, the method may further include the following: determining the weight, similarity, and frequency of occurrence of the related technical elements; calculating the relevance parameters of the related technical elements based on their weight, similarity, and frequency of occurrence; selecting the technical element with the highest relevance parameter from the related technical elements based on the relevance parameter, and generating the prompt keyword for the current round based on this technical element. This allows for the generation of prompt keywords for the current round that yield relatively better results.

[0225] In some embodiments, the above-mentioned method of determining the prompt words for the current round based on the response text of the previous round and the target technical element may further include the following: querying the target knowledge graph based on the response text of the current round to determine whether there is a related next technical element; and ending the multi-round questioning if it is determined that there is no related next technical element.

[0226] In practice, based on the response text of the current round, the target knowledge graph can be retrieved using appropriate graph algorithms to detect whether a related next technical element exists. If no related technical element is found, it can be determined that further in-depth questioning based on the target knowledge graph is not possible. At this point, the multi-round questioning can be terminated. Based on the multiple response texts obtained from the multi-round questioning, a corresponding patent summary can be generated.

[0227] In some cases, the system can also receive and re-question based on a previously unaddressed technical element re-entered by the user. The responses to the previous question and the responses to the re-question are then combined to generate the final patent summary.

[0228] Based on the above embodiments, the timing for ending multiple rounds of questioning can be accurately determined according to the target knowledge graph, and the multiple rounds of questioning can be ended in a timely manner.

[0229] In some embodiments, after using the prompt words of the current round to ask questions to the patent summary model based on the patent content of the target patent document and obtaining the response text of the current round, the method may further include the following: performing semantic matching by using relevant embodiments in the target patent document to detect whether the response text of the current round is a response text that meets the requirements; and ending the multi-round questioning if it is determined that the response text of the current round is a response text that meets the requirements.

[0230] In practice, the target patent document can be searched based on the prompt words of the current round to determine the relevant embodiments; then, semantic recognition can be performed on the relevant embodiments to obtain the corresponding semantic recognition results; and the semantic matching degree between the semantic recognition results and the response text of the current round can be calculated.

[0231] Based on the semantic matching degree, when it is determined that the semantic recognition result matches the response text of the current round with a high degree of matching, it can be determined that the response text of the current round meets the requirements; then it can be determined that the required response text has been obtained, and the multi-round questioning ends.

[0232] Conversely, based on semantic matching degree, if it is determined that the semantic recognition result has a low matching degree with the response text of the current round, the next round of questioning can continue until a response text that meets the requirements is obtained.

[0233] Based on the above embodiments, the timing for ending multiple rounds of questioning can be accurately determined according to the relevant embodiments in the target patent document, and the multiple rounds of questioning can be ended in a timely manner.

[0234] In some embodiments, after using the prompt words of the current round to ask questions to the patent summary model regarding the patent content of the target patent document and obtaining the response text of the current round, the method may further include: detecting whether the response text of the current round is empty; if it is determined that the response text of the current round is empty, it can be determined that it is impossible to continue asking questions to the patent summary model and obtain more response texts, and at this time, the multi-round questioning can be terminated in a timely manner.

[0235] Furthermore, after using the prompt words of the current round to ask the patent summary model questions based on the patent content of the target patent document and obtaining the response text of the current round, the method may further include: detecting whether the response text of the current round is semantically duplicated with the response texts of previous rounds; if it is determined that the response text of the current round is semantically duplicated with the response texts of previous rounds, it can be determined that continuing to ask questions to the patent summary model will not yield a response text containing new content, and at this time, the multi-round questioning can be terminated in a timely manner.

[0236] In some embodiments, after the completion of multiple rounds of questioning, the method further includes: determining technical problem text, technical means text, and technical effect text based on the response texts of the multiple rounds of questioning; and combining the technical problem text, technical means text, and technical effect text to obtain the patent summary of the target patent document.

[0237] In practice, semantic recognition can be performed on the response texts of multiple rounds of questions to obtain semantic recognition results for each response text. Based on the semantic recognition results, technical issue texts, technical means texts, and technical effect texts can be selected from the response texts of multiple rounds of questions.

[0238] After obtaining the semantic recognition results, multiple response texts related to technical issues can be selected from the response texts of multi-round questions as the first type of response text, response texts related to technical means can be selected as the second type of response text, and response texts related to technical effects can be selected as the third type of response text. Then, the first type of response text is used in combination to generate the technical issue text, the second type of response text is used in combination to generate the technical means text, and the third type of response text is used in combination to generate the technical effect text.

[0239] In practice, the technical problem text, technical means text, and technical effect text can be sequentially assembled to obtain the patent summary of the target patent document.

[0240] Based on the above embodiments, a relatively complete patent summary can be generated by combining the response texts of multiple rounds of questions.

[0241] In some embodiments, Figure 6 This is a schematic flowchart illustrating the process of training a patent summary model using the patent summary generation method of this application's embodiments in a scenario example. Figure 6 As shown, the method, in its specific implementation, may include the following:

[0242] S601: Construct an initial model based on a large language model;

[0243] S602: Using the first patent sample data, train the initial model to obtain the first intermediate model that meets the requirements;

[0244] S603: Using the second patent sample data, train the first intermediate model to obtain a second intermediate model that meets the requirements; wherein, the second patent sample data is labeled with a corresponding patent summary;

[0245] S604: Fine-tune the second intermediate model and add a preset loop guidance mechanism to obtain a patent summary model that meets the requirements.

[0246] In practice, a general large language model can be used as the initial model.

[0247] In practice, sample patent documents can be obtained first as the first patent sample data; then, the first patent sample data can be used to train an initial model to obtain a large language model applicable to the patent document field, which serves as the first intermediate model that meets the requirements.

[0248] Figure 7 Another flowchart illustrating the process of training a patent summary model using a patent summary generation method that applies an embodiment of this application in a scenario example is shown.

[0249] For specific implementation, please refer to Figure 7 As shown, sample patent documents can be obtained first, and the corresponding patent summary can be marked on the sample patents to obtain the corresponding second patent sample data. Then, the second patent sample data can be used to train the first intermediate model to obtain a large language model that can automatically process patent documents and extract the general patent summary based on the inventive points in the patent documents, which serves as the second intermediate model that meets the requirements.

[0250] In practice, the model can be fine-tuned based on the second intermediate model and the relevant rules for generating the patent summary; and a preset loop guidance mechanism can be added to the model.

[0251] The aforementioned preset cyclic guidance mechanism may include multiple preset cyclic guidance mechanisms corresponding to different summary types, so as to obtain a large language model that supports multi-round progressive questioning and can automatically generate a patent summary with high accuracy, as a qualified patent summary model.

[0252] Specifically, the aforementioned pre-defined cyclical guidance mechanisms may include: easy-to-understand guidance mechanisms, core summary guidance mechanisms, comprehensive summary guidance mechanisms, higher-level summary guidance mechanisms, and personalized guidance mechanisms.

[0253] Based on the above embodiments, a patent summary model with good performance and high accuracy can be trained that is suitable for processing patent documents to obtain patent summaries.

[0254] In some embodiments, the patent summary model described above includes at least a core summary guidance mechanism. Correspondingly, during training, the patent summary model also learns and masters the processing logic of determining appropriate technical elements based on their weighting to ask questions and uncover the most core inventive points.

[0255] Figure 8 This diagram illustrates a process flow diagram of a patent summary generation method applying embodiments of this application in a scenario example, utilizing a core summary-based guiding mechanism based on a patent summary model. When the type parameter is of the core summary type, such as... Figure 8 As shown, the above-mentioned questioning of the patent summary model based on the patent content of the target patent document according to the target knowledge graph can specifically include the following: First, based on the target knowledge graph, basic question sentences (e.g., what are the core technical problems, technical means, and technical effects) are generated using basic prompt words (e.g., technical problems, technical means, and technical effects); and the basic question sentences are input into the patent summary model to ask questions and obtain basic response text.

[0256] Based on the textual content of the technical issues in the basic response text, and according to the technical issue knowledge graph or target knowledge graph corresponding to the background technology module (corresponding to the background technology in the specification), the technical elements with the highest weight are selected by statistically analyzing the number of edges between nodes corresponding to relevant technical elements to generate the optimal prompt words for the technical issues. These prompt words are then used to generate a question about the technical issues. This question is then input into the patent summary model for questioning, resulting in the response text for the core technical issues. Following this method, based on the core summary guidance mechanism, multiple rounds of questioning are conducted according to the technical issue knowledge graph to ultimately obtain the technical issue text that meets the requirements for the core technical issues.

[0257] Figure 9 This diagram illustrates a detailed processing flow of a patent summary generation method applying embodiments of this application in a scenario example, utilizing a core generalization-based guiding mechanism based on a patent summary model. For example, as... Figure 9 As shown, based on the target knowledge graph and the core summary guidance mechanism, the optimal prompt word "on-site" can be determined through multiple rounds of questioning; then, based on this prompt word, a question sentence is generated to ask the patent summary model, so as to finally obtain the core technical problem as "high cost".

[0258] Simultaneously, based on the textual content of technical means involved in the basic response text, and according to the technical means knowledge graph corresponding to the specification content module (corresponding to the invention content in the specification) and the embodiment module (corresponding to the specific implementation methods in the specification), the technical elements with the highest weight are selected by statistically analyzing the number of edges corresponding to the nodes of relevant technical elements to generate the optimal prompt words about technical means. These prompt words are then used to generate a question about the technical means. This question is then input into the patent summary model for questioning, resulting in a response text about the core technical means. Following this method, based on the core summary guidance mechanism, multiple rounds of questioning are conducted according to the technical means knowledge graph, ultimately yielding a technical means text that meets the requirements regarding the core technical means.

[0259] Meanwhile, based on the text content related to technical effects in the basic response text, and according to the technical effect knowledge graph corresponding to the specification content module and the embodiment module, the technical elements with the largest weight are selected by statistically analyzing the number of edges corresponding to the nodes of the relevant technical elements to generate the optimal prompt words about the technical effects; and using these prompt words, a question about the technical effects is generated; and then the question is input into the patent summary model to ask questions, thereby obtaining the response text about the core technical effects.

[0260] Following the above approach, based on a core summary-driven guidance mechanism, multiple rounds of questioning are conducted using a technical means knowledge graph to ultimately obtain a technical effect text that meets the requirements regarding the core technical effects.

[0261] Finally, by combining the above-mentioned technical problem text, technical means text, and technical effect text, a corresponding patent summary outlining the core inventive points was generated.

[0262] Figure 10 This diagram illustrates a detailed processing flow of a patent summary generation method applying embodiments of this application in a scenario example, utilizing a core generalization-based guiding mechanism based on a patent summary model. For example, as... Figure 10 As shown, a patent summary in the form of a one-line offer can be generated according to the corresponding requirements, which includes text on technical issues, technical means, and technical effects.

[0263] This allows the use of a patent summary model to generate a patent summary that includes the core inventive points and highlights key information.

[0264] In some embodiments, the patent summary model described above can be equipped with an easy-to-understand guidance mechanism. Accordingly, during the training process, the patent summary model will also learn and master the mapping relationship between patent expressions and general expressions (e.g., the correspondence between patent terms and their corresponding general terms), as well as the processing logic for mining and generating more easily understandable patent summaries using this mapping relationship.

[0265] Figure 11 This diagram illustrates a process flow diagram of a patent summary generation method applying embodiments of this application in a scenario example, utilizing a patent summary model based on an easy-to-understand guiding mechanism. When the type parameter is easy to understand, such as... Figure 11 As shown, the above-mentioned questioning of the patent summary model based on the patent content of the target patent document according to the target knowledge graph can specifically include the following: First, based on the target knowledge graph, select the target technical elements that meet the requirements; based on the target technical elements, generate corresponding prompt words; wherein, the above prompt words are patent terms based on the patent description, such as elastic element, elastic component, mechanical component, etc.

[0266] Then, based on the above prompts, generate corresponding questions based on the patent description; and input the questions into the patent summary model to ask questions; the patent summary model processes the questions, generates corresponding response text, and can also replace the patent description in the response text with the corresponding general description according to the mapping relationship between the patent description and the general description. For example, replace the elastic element with a spring, and output response text based on the general description.

[0267] By repeating the above approach and using an easy-to-understand guiding mechanism, multiple rounds of questions can be posed to the patent summary model to obtain multiple response texts based on common expressions. These multiple response texts can then be combined to generate a corresponding patent summary based on common expressions.

[0268] This allows the use of patent summary models to generate patent summaries that are generally applicable and easy for users to understand, based on an easy-to-understand guiding mechanism.

[0269] In some embodiments, the patent summary model described above may be equipped with a comprehensive guidance mechanism. Accordingly, during the training process, the patent summary model will also learn and master the processing logic for mining technical elements involving multiple inventive points from the knowledge graph, and generating a patent summary that simultaneously covers the multiple inventive points involved using the aforementioned technical elements.

[0270] When the type parameter is comprehensive and general, the above-mentioned questioning of the patent summary model based on the patent content of the target patent document according to the target knowledge graph can specifically include the following: based on the target knowledge graph and the comprehensive and general guidance mechanism, the patent summary model is used to ask questions about each embodiment in the multiple embodiments of the embodiment module of the target patent document to obtain the response text for each embodiment; then the response texts of each embodiment are combined and the relevant inventive points are sorted out to finally obtain a patent summary that contains multiple inventive points appearing in the target patent document.

[0271] This allows the use of a patent summary model with a comprehensive overview-based guiding mechanism to generate a patent summary that fully covers multiple inventive points involved in the target patent document.

[0272] In some embodiments, the patent summary model described above can be supplemented with a higher-level generalization guidance mechanism. Accordingly, during the training process, the patent summary model will also learn and master the mapping relationship between lower-level and higher-level statements, as well as the processing logic of using this mapping relationship to mine and generate a higher-level patent summary that is more comprehensive than the basic solution.

[0273] When the type parameter is a higher-level generalization, the above-mentioned questioning of the patent summary model based on the patent content of the target patent document according to the target knowledge graph can specifically include the following: First, based on the target knowledge graph, select the target technical elements that meet the requirements; generate corresponding prompt words based on the target technical elements; then generate corresponding question sentences based on the prompt words and ask the patent summary model; the patent summary model processes the question sentences, generates corresponding response text, and at the same time, it can detect the lower-level expressions in the response text according to the mapping relationship between the lower-level expressions and the higher-level expressions, and replace the lower-level expressions with the corresponding higher-level expressions to obtain the adjusted response text; then, based on the adjusted response text, perform generalization and higher-level processing to obtain the higher-level generalized response text.

[0274] By repeating the above-mentioned approach, and using the guidance mechanism based on higher-level generalizations, multiple rounds of questions can be asked on the patent summary model to obtain multiple response texts based on higher-level generalizations. These multiple response texts based on higher-level generalizations can then be combined to generate a corresponding patent summary based on higher-level generalizations.

[0275] Thus, a relatively higher-level, more comprehensive patent summary can be generated by using the patent summary model based on the higher-level generalization guidance mechanism.

[0276] In some embodiments, the patent summary model described above may be equipped with a personalized guidance mechanism. Accordingly, during the training process, the patent summary model will also learn and master some personalized mapping relationships and related processing logic.

[0277] Specifically, for example, the aforementioned patent summary model can also be linked to user-defined custom rules (e.g., including customized expression rules and customized technical dictionaries).

[0278] When the type parameter is customized, after generating the response text, the patent summary model can modify and adjust the expression and technical terms used in the response text according to the customized expression rules and customized technical dictionary to obtain the modified and adjusted response text; then, by combining the modified and adjusted response text, a patent summary that meets the user's personalized requirements can be obtained.

[0279] This allows the use of a patent summary model with a personalized guidance mechanism to generate patent summaries that meet the user's individual requirements.

[0280] In some embodiments, depending on the specific circumstances and processing needs, a patent summary model can be used, along with multiple preset cyclical guidance mechanisms, to ask questions and obtain the corresponding patent summary. This can satisfy more diverse and complex personalized needs.

[0281] Second aspect of the embodiments

[0282] An embodiment of the second aspect of this application provides a patent summary generation apparatus, the patent summary generation apparatus comprising:

[0283] Memory, which is used to store computer programs;

[0284] A processor that, when executing the computer program, implements the method for generating any of the patent outlines provided in the first aspect of the embodiments of this application.

[0285] Figure 12 This is a schematic diagram of the structure of a patent summary generation device according to an embodiment of this application. For example... Figure 12 As shown, the patent summary generation device 1200 may include a processor 1210 and a memory 1220; the memory 1220 is coupled to the processor 1210. It is worth noting that this figure is exemplary; other types of structures may be used to supplement or replace this structure to achieve telecommunications functions or other functions.

[0286] Specifically, the processor 1210 can be used to acquire a target knowledge graph corresponding to the target patent document; based on the target knowledge graph, to ask questions about the patent summary model for the patent content of the target patent document, thereby obtaining a patent summary of the target patent document; wherein, the patent summary model includes a large language model;

[0287] The memory 1220 can be used to store the corresponding instruction program.

[0288] In this embodiment, the processor 1210 can be implemented in any suitable manner. For example, the processor can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers, etc. This application embodiment is not limited to any particular type.

[0289] In this embodiment, the memory 1220 may include multiple layers. In a digital system, anything that can store binary data can be a memory. In an integrated circuit, a circuit with storage function but no physical form is also called a memory, such as RAM, FIFO, etc. In a system, a storage device with a physical form is also called a memory, such as a memory stick, TF card, etc.

[0290] In addition, the patent summary generation device 1200 may also include other structural modules such as a communication module, an input / output unit, a display, and a power supply.

[0291] This application also provides a computer-readable program, wherein when the program is executed, the program causes the computer to perform a method for generating any patent summary as described in the embodiments of the first aspect of this application.

[0292] This application also provides a computer-readable storage medium storing a computer program that causes a computer to execute the method for generating any patent summary as described in the embodiments of the first aspect of this application.

[0293] This application also provides a computer program product, wherein when the computer program product is executed by a processor, it causes the computer to perform the patent summary generation method described in the first aspect of this application.

[0294] The apparatus and methods described above in the embodiments of this application can be implemented in hardware or in combination with software. This application relates to a computer-readable program that, when executed by a logic component, enables the logic component to implement the apparatus or constituent parts described above, or to implement the various methods or steps described above.

[0295] This application also relates to storage media for storing the above programs, such as hard disks, magnetic disks, optical disks, DVDs, flash memory, etc.

[0296] Third aspect of the embodiments

[0297] An embodiment of the third aspect of this application provides a patent summary generation system, which includes the patent summary generation device described in the second aspect of this application.

[0298] Based on the aforementioned patent summary generation system, by introducing and using a knowledge graph corresponding to the patent document as an aid during the question-and-answer dialogue with the patent summary model, appropriate prompt words can be determined relatively efficiently and accurately. Based on these prompt words, the model can be automatically and deeply questioned multiple times to obtain the corresponding patent summary.

[0299] Fourth aspect of the embodiment

[0300] Figure 13 This is a schematic diagram of a device for generating a patent summary according to an embodiment of this application. Figure 13 As shown, the patent summary generation apparatus may specifically include the following modules:

[0301] The acquisition module 1310 can be used to acquire the target knowledge graph corresponding to the target patent document.

[0302] The processing module 1320 is specifically used to ask questions about the patent summary model based on the target knowledge graph and the patent content of the target patent document to obtain the patent summary of the target patent document; wherein, the patent summary model includes a large language model.

[0303] The patent summary generation device described above can efficiently and accurately determine appropriate prompts by introducing and using a knowledge graph corresponding to the patent document during the question-and-answer dialogue with the patent summary model. Based on these prompts, it can automatically and deeply ask the model multiple rounds of questions to obtain the corresponding patent summary.

[0304] It should be noted that the limitations on each step involved in this application are not considered as limiting the order of steps, provided that they do not affect the implementation of the specific solution. The steps listed first can be performed first, later, or even simultaneously. As long as the solution can be implemented, they should be considered to fall within the scope of protection of this application.

[0305] The present application has been described above with reference to specific embodiments. However, those skilled in the art should understand that these descriptions are exemplary and not intended to limit the scope of protection of the present application. Those skilled in the art can make various modifications and variations to the present application based on its spirit and principles, and these modifications and variations are also within the scope of the present application.

Claims

1. A method of generating a patent abstract, characterized by, The method comprises: obtaining a target knowledge graph corresponding to a target patent file; According to the target knowledge graph, the patent summary model is questioned for the patent content of the target patent file, and the patent summary of the target patent file is obtained. The patent summary model comprises a large language model.

2. The method of claim 1, wherein, The target knowledge graph at least comprises technical elements in the target patent file and the correlation between the technical elements.

3. The method of claim 1, wherein, Obtaining a target knowledge graph corresponding to a target patent file comprises: According to the patent text features of the target patent file, a target knowledge graph about the target patent file is generated; The patent text features include patent text structure features and patent text vocabulary features.

4. The method according to claim 2 or 3, characterized in that, Obtaining a target knowledge graph corresponding to a target patent file comprises: Entity recognition is performed on the target patent file to determine the entity objects in the target patent file; wherein the entity objects at least include technical elements; Through relationship extraction on the target patent file, the correlation between the entity objects is obtained; According to the entity objects and the correlation between the entity objects, a target knowledge graph corresponding to the target patent file is constructed.

5. The method of claim 4, wherein, The technical elements include technical means elements, technical problem elements and technical effect elements.

6. The method of claim 3, wherein, According to the patent text features, a target knowledge graph about the target patent file is generated, comprising: According to the patent text features, the target patent file is divided into multiple modules; wherein the multiple modules include background technology module, specification content module, embodiment module; Entity recognition and relationship extraction are performed on the multiple modules respectively to obtain the entity objects of the multiple modules and the correlation between the entity objects; According to the entity objects of the multiple modules and the correlation between the entity objects, the corresponding technical problem knowledge graph, technical means knowledge graph and technical effect knowledge graph are constructed.

7. The method of claim 1, wherein, According to the target knowledge graph, the patent summary model is questioned for the patent content of the target patent file, and the patent summary of the target patent file is obtained, comprising: According to the target knowledge graph, the target technical elements meeting the requirements are screened out; According to the target technical elements screened out, the corresponding prompt words are generated; Using the prompt words, the patent summary model is questioned for the patent content of the target patent file, and the corresponding reply text is obtained; According to the reply text, the patent summary of the target patent file is generated.

8. The method of claim 7, wherein, According to the target knowledge graph, the target technical elements meeting the requirements are screened out, comprising: At least one target technical element is screened out from different knowledge graphs in the target knowledge graph as the target technical element meeting the requirements; The target knowledge graph comprises a technical problem knowledge graph, a technical means knowledge graph and a technical effect knowledge graph.

9. The method of claim 7, wherein, According to the target knowledge graph, the target technical elements meeting the requirements are screened out, comprising: Obtaining a summary type parameter; According to the summary type parameter and the target knowledge graph, the target technical elements meeting the requirements are screened out; The summary type parameter includes one or more of the following: easy-to-understand type, core summary type, comprehensive summary type, upper summary type, and personalized customization type.

10. The method of claim 7, wherein, The type of the prompt word includes at least one of the following: Text prompt, image prompt, and sound prompt.

11. The method of claim 7, wherein, According to the reply text, a patent summary of the target patent file is generated, including: According to the summary template, the patent summary of the target patent file is generated using the reply text.

12. The method of claim 11, wherein, The summary template includes at least one of the following: Text content of the patent summary in the order of technical problems, technical means, and technical effects; Text length of the text content of the generated patent summary; Language units of the text content of the generated patent summary; Language style of the text content of the generated patent summary.

13. The method of claim 7, wherein, Using the prompt word, the patent summary model is questioned for the patent content of the target patent file, and the patent summary of the target patent file is obtained, including: Using the prompt word, the patent summary model is questioned for the patent content of the target patent file, and the patent summary of the target patent file is obtained.

14. The method of claim 13, wherein, Using the prompt word, the patent summary model is questioned for the patent content of the target patent file, including: Each round of questioning in the multi-round questioning of the patent summary model for the patent content of the target patent file includes: Obtaining the reply text of the previous round; According to the reply text of the previous round and the target technical element, the prompt word of the current round is determined; Using the prompt word of the current round, the patent summary model is questioned for the patent content of the target patent file, and the reply text of the current round is obtained.

15. The method of claim 14, wherein, According to the reply text of the previous round and the target technical element, the prompt word of the current round is determined, including: Extracting the subject word from the reply text of the previous round; According to the subject word and the target technical element of the previous round, the target knowledge graph is queried to determine the technical element with a required similarity; According to the technical element with a required similarity, the prompt word of the current round is generated.

16. The method of claim 15, wherein, In the case where the technical element with a required similarity includes a plurality of technical elements with a required similarity, the method further includes: Determining the technical element with the largest similarity from the plurality of technical elements with a required similarity; and generating the prompt word of the current round according to the technical element.

17. The method of claim 14, wherein, According to the reply text of the previous round and the target technical element, the prompt word of the current round is determined, further including: Extracting the subject word from the reply text of the previous round; According to the subject word and the target technical element of the previous round, the target knowledge graph is queried to determine the associated technical elements; According to the weight of the technical elements, the technical element with the largest weight ratio is determined from the associated technical elements; and the prompt word of the current round is generated according to the technical element.

18. The method of claim 14, wherein, According to the reply text of the previous round and the target technical element, the prompt word of the current round is determined, further including: According to the reply text of the current round, the target knowledge graph is queried to determine whether there is an associated next technical element; In the case where it is determined that there is no associated next technical element, the multi-round questioning is ended.

19. The method of claim 14, wherein, After the current round of question and answer is performed on the patent content of the target patent document by using the prompt word of the current round, the method further comprises: By using the related embodiments in the target patent document for semantic matching, it is detected whether the current round of answer text is a required answer text; In the case where the current round of answer text is determined to be a required answer text, the multi-round question and answer is ended.

20. The method of claim 18 or 19, wherein, After the multi-round question and answer is ended, the method further comprises: According to the answer texts of the multi-round question and answer that have been performed, technical problem text, technical means text and technical effect text are determined; The technical problem text, the technical means text and the technical effect text are combined to obtain the patent summary of the target patent document.

21. The method of claim 1, wherein, The method further comprises: An initial model based on a large language model is constructed; The initial model is trained by using first patent sample data to obtain a first intermediate model that meets the requirements; The first intermediate model is trained by using second patent sample data to obtain a second intermediate model that meets the requirements; wherein the second patent sample data is labeled with a corresponding patent summary; The second intermediate model is fine-tuned and a preset loop guiding mechanism is added to obtain a required patent summary model.

22. A patent abstract generating apparatus characterized by comprising: The patent summary generation device comprises: a memory storing a computer program; and a processor that, when executing the computer program, implements the patent summary generation method of any one of claims 1 to 21.

23. A patent summary generation system, characterized in that, the patent summary generation system comprises the patent summary generation device of claim 22.

Images