Technique document optimization method and device based on language model

Abstract

Embodiments of the present application provide a technical document optimization method and device based on a language model. The method comprises: parsing a body of a first technical document to obtain a structured abstract of the body of the first technical document, the body of the first technical document comprising an innovation scheme; mapping the structured abstract to a technical document optimization strategy, the technical document optimization strategy comprising an optimization strategy corresponding to the structured abstract; and generating a second technical document according to the optimization strategy corresponding to the structured abstract, the body of the second technical document comprising the original meaning of the innovation scheme, and the structure and semantics of the body of the second technical document being aligned. Embodiments of the present application use a dual-channel of a strategy engine and a neural network, are uniformly scheduled by a strategy library, can enhance the readability and understandability of the technical document on the basis of retaining the original innovation scheme, and can effectively improve the efficiency of subsequent business processing.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and more specifically, to a method and apparatus for optimizing technical documents based on a language model. Background Technology

[0002] Technical documents are the textual carriers of technological innovation solutions, and the quality of technical documents is a key indicator for measuring the efficiency of technology disclosure.

[0003] A common problem with technical documentation is that it's "written for oneself, not for the reader." Common issues include disorganized and difficult-to-read layouts, disjointed steps or chapters, excessive use of abbreviations that increase the learning curve, lack of examples, omissions, typos, or excessively long paragraphs. Traditionally, these problems are addressed through manual correction, which is time-consuming and costly. Alternatively, using generic, large-scale models to automatically generate optimized documentation may result in fabricated or irrelevant content, leading to redundancy and failing to capture the user's true points of confusion.

[0004] For example, the inventor verbally explains or revises the technical disclosure document through one-on-one meetings, phone calls, or emails by the agent or technician. Because this manual revision heavily relies on the agent's personal professional background and experience, both time and manpower costs are high. Furthermore, information may be misinterpreted when it is transmitted from the inventor to the agent, and experienced agents find it difficult to scale up. The method based on a general large model involves directly inputting the original technical disclosure document into the language model and freely rewriting it according to general prompts. Documents generated in this way lack a patent layout strategy, making it difficult to guarantee the quality, stability, and compliance of document optimization, and it is also difficult to handle complex technical solutions. Summary of the Invention

[0005] This application provides a method and apparatus for optimizing technical documents based on a language model. It can complete the alignment of document structure and semantics through a dual-channel approach of strategy engine and language model while retaining the original innovative solutions. This can enhance the readability and comprehensibility of technical documents while preserving the original innovative solutions, and can effectively improve the efficiency of subsequent business processing.

[0006] Firstly, a language model-based method for optimizing technical documents is provided. This method includes: parsing the main text of a first technical document to obtain a structured summary of the main text, the main text of which includes an innovative solution; mapping the structured summary to a technical document optimization strategy, the optimization strategy including an optimization strategy corresponding to the structured summary; and generating a second technical document based on the optimization strategy corresponding to the structured summary, the main text of which includes the original meaning of the innovative solution, and the structure and semantics of the main text of the second technical document are aligned.

[0007] In some embodiments, mapping the structured summary to a technical document optimization strategy may include: mapping the structured summary to the technical document optimization strategy to form a decision table for the first technical document, the decision table including optimization strategy entries in the technical document optimization strategy corresponding to the structured summary; generating a second technical document based on the optimization strategy corresponding to the structured summary may include: generating the second technical document based on the optimization strategy entries included in the decision table.

[0008] In some embodiments, the technical document optimization strategy includes a hard constraint strategy and / or a soft preference strategy; wherein the soft preference strategy includes one or more of the following: a derivation strategy for embodiments of the innovative solution, a keyword explanation or overarching strategy for the innovative solution, a sub-solution decomposition of the innovative solution into a complex solution, a problem list generation strategy for the innovative solution, a template and format for the technical document, a strategy for creating diagrams for the innovative solution, a personal style setting strategy for the technical document, and a hierarchical progressive description method for the innovative solution from abstract to concrete or from general to specific; the hard constraint strategy includes format specifications and standard clauses.

[0009] In some embodiments, the method may further include: obtaining a third technical document, which is a revised version of the second technical document; using the difference between the third technical document and the decision table as feedback entries to precipitate new strategy entries to update the technical document optimization strategy; and using the feedback entries as preference alignment data to fine-tune the language model.

[0010] Optionally, the method may further include: using the third technical document as the first technical document for the next input to the language model; or using the second technical document obtained from the i-th correction as the first technical document for the (i+1)-th input to the language model, where i is a positive integer.

[0011] In some embodiments, the decision table for mapping the structured summary to the technical document optimization strategy to form the first technical document may include: matching the structured summary to the hard constraint strategy based on a policy engine to obtain a first matching result; matching the structured summary to the soft preference strategy based on a neural network to obtain a second matching result; and performing decision fusion on the first matching result and the second matching result to generate the decision table, wherein the decision fusion prioritizes the first matching result over the second matching result.

[0012] In some embodiments, the first technical document is a first technical disclosure, and the second technical document is an optimized second technical disclosure; wherein the structured summary includes at least one of the following: the type of the innovation, the hierarchy of the innovation, and the number of entities of the innovation.

[0013] In some embodiments, obtaining a structured summary of the first technical document by parsing its main text may include: parsing the main text of the first technical disclosure document using a layout language model or optical character recognition technology to identify the positions of paragraphs, tables, and / or figures; extracting the technical problem to be solved by the innovative solution, the elements, parameter indicators, and / or figure numbers of the innovative solution using a conditional random field; and using a multi-label classifier to predict type labels and / or hierarchical labels for the first technical disclosure document to obtain the structured summary.

[0014] In some embodiments, generating a second technical document based on the decision table may include: selecting a corresponding writing template according to the applicable template number in the decision table; generating, in parallel by chapter, various components of the second technical disclosure document according to the structured elements of the decision table and the main text of the first technical disclosure document, wherein each component of the second technical disclosure document includes the prior art and technical problems closest to the innovative solution, an overview of the innovative solution, specific embodiments of the innovative solution, and a list of problems; performing constraint decoding on each component of the second technical disclosure document to satisfy the list of hard constraints in the decision table; and performing hard constraint verification on each component of the second technical disclosure document, and rewriting the parts that fail according to the corresponding strategy entries in the decision table.

[0015] In some embodiments, the process of generating the components of the second technical disclosure document in parallel by chapter based on the structured elements of the decision table and the main body of the first technical disclosure document may include: determining the distinguishing features between the innovative solution and the prior art as a query vector; recalling historical extended embodiment fragments that are adjacent to the query vector in a vector library; generating extended embodiments based on the distinguishing features, the historical extended embodiment fragments, and the extension type preferences in the decision table, according to parametric extension, combined extension, and / or boundary condition extension; and adding the extended embodiments to the issue list.

[0016] In some embodiments, the parallel generation of the various components of the second technical disclosure document by chapter according to the structured elements of the decision table and the main text of the first technical disclosure document may further include: determining that the source drawings in the first technical disclosure document are to be retained, and retaining them inline into the second technical disclosure document according to the vector format in the technical document optimization strategy; determining that the source drawings in the first technical disclosure document are to be redrawn, obtaining editable vector graphics by rendering the structured code corresponding to the source drawings and embedding them into the second technical disclosure document and executing the drawing creation strategy constraints in the technical document optimization strategy; and / or generating editable graphics corresponding to specific embodiments of the second technical disclosure document and embedding them into the second technical disclosure document.

[0017] In some embodiments, the language model is a language model based on a decoder-integrated self-attention mechanism; the language model sequentially includes a tokenizer, an embedding layer, N decoding blocks, an output linear layer, and a normalized output layer, where N is an integer greater than or equal to 1; each decoding block sequentially includes a masked multi-head self-attention sub-layer, a first normalization layer, and a second normalization layer; the first technical document is input into the language model, and the second technical document is output based on the language model; the tokenizer outputs a decision table and structured elements of the first technical document based on the first technical document and the technical document optimization strategy; the second technical document is generated by passing through the embedding layer, the N decoding blocks, the output linear layer, and the normalized output layer, and the normalized output layer outputs the parallel generated word sequence of each chapter of the second technical document in the normalized output layer. Optionally, N is greater than 80 and less than 120.

[0018] In some embodiments, the output of the normalized output layer can be used as the input to the tokenizer. This enables self-correction and / or sampling diversity, reducing term drift or potential inconsistencies in content.

[0019] In a second aspect, a language model-based technical document optimization device is provided, the device including a processor and a memory, the processor, the memory and a network interface being interconnected via a communication bus; the memory is used to store computer program instructions; the processor is used to read the computer program instructions stored in the memory to execute the method in any possible implementation of the first aspect above.

[0020] In some embodiments, the device may further include a network interface for receiving the first technical document and / or sending the second technical document.

[0021] Thirdly, a computer-readable storage medium is provided that stores a computer program, which, when executed, implements the method in any possible implementation of the first aspect described above.

[0022] In this embodiment, a second technical document with structured and semantic alignment is generated through a dual-channel strategy engine and language model, uniformly scheduled by a strategy library. This process preserves the original intent of the innovative solution in the original technical document while proactively meeting the hard constraints of the patent application, providing decision traceability, and supporting synchronous updates of the strategy library and language model based on user feedback. In summary, this reduces communication costs and information transmission discrepancies between the agent and the inventor, avoids unauthorized rewriting and missing elements issues arising from freely generated large models, and thus improves the efficiency of technical document optimization and subsequent business processing. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic flowchart illustrating the process of a language model-based technical document optimization method according to an embodiment of this application.

[0025] Figure 2 This is a schematic flowchart illustrating the process of optimizing a technical disclosure method according to an embodiment of this application.

[0026] Figure 3 This is a schematic flowchart illustrating a technical disclosure optimization method based on a decision table according to an embodiment of this application.

[0027] Figure 4 This is a schematic flowchart illustrating the process of optimizing a technical disclosure method according to an embodiment of this application.

[0028] Figure 5 This is a schematic diagram of the dual-channel parallel decision fusion structure according to an embodiment of this application.

[0029] Figure 6 This is a schematic diagram of the structure generated from the accompanying drawings of one embodiment of this application.

[0030] Figure 7 This is a schematic diagram of a language model structure based on a decoder-integrated self-attention mechanism according to an embodiment of this application.

[0031] Figure 8 This is a structural block diagram of a technical document optimization device according to an embodiment of this application. Detailed Implementation

[0032] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0033] It should be understood that the terms "first" and "second" in the embodiments of this application are for distinction only and should not constitute any limitation on this application. It should also be understood that in the various embodiments of this application, the sequence number of each process does not imply the order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0034] It should also be noted that "and / or" describes the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the related objects before and after it are in an "or" relationship.

[0035] The technical solution of this application can be applied to the processing of technical documents in research institutions, the compilation of technical disclosure documents by agencies, AI-assisted patent drafting, the technical mining of enterprise R&D documents, and other scenarios that transform unstructured technical texts into structured documents according to policy constraints. Typical application scenarios include, but are not limited to, invention patent technical disclosure documents, utility model technical disclosure documents, PCT international application technical disclosure documents, research reports, patent application documents, and papers.

[0036] Technical documents include, but are not limited to, technical disclosure texts with innovative solutions that address problems existing in the prior art; the forms of these technical documents include, but are not limited to, technical disclosure documents, R&D reports, technical solution documents, patent application documents, and experimental records.

[0037] Language models can include, but are not limited to, language models based on decoder-only Transformer (DET) or GPT, Claude, Qwen, and other language models.

[0038] Figure 1 This is a schematic flowchart of a language model-based technical document optimization method according to an embodiment of this application. Figure 1 The method is executed by the technical documentation optimization device.

[0039] 101. Parse the main text of the first technical document to obtain a structured summary of the main text of the first technical document, which includes innovative solutions.

[0040] 102. Map the structured summary to a technical document optimization strategy, which includes optimization strategies corresponding to the structured summary.

[0041] 103. Generate a second technical document according to the optimization strategy corresponding to the structured summary. The main body of the second technical document includes the original meaning of the innovative solution, and the structure and semantics of the main body of the second technical document are aligned.

[0042] In some embodiments, the first technical document may be provided by the user and may include at least an innovative solution to the technical problem. In some embodiments, the first technical document may also have defects in terms of content and / or format.

[0043] In some embodiments, the second technical document includes the original intent of the innovative solution in the first technical document (which may be different expressions of the same technology). In other words, the innovative solution in the second technical document is essentially the same as the innovative solution in the first technical document (or both express the same technology). The text structure and semantic alignment of the second technical document are as follows. Structure and semantic alignment includes, but are not limited to: key terms with specific explanations, tight contextual connection, fluent sentences, consistent terminology for the same concept throughout the text, and / or the inclusion of easily understandable figures in appropriate locations (such as supplementing a solution with a figure for greater clarity).

[0044] In some embodiments, in step 101, a text extraction profile of the first technical document can be parsed, the profile including a structured summary of the innovative solution; in step 102, the profile is mapped to a technical document optimization strategy, the optimization strategy including an optimization strategy corresponding to the profile; and in step 103, a second technical document is generated based on the optimization strategy corresponding to the profile. The structure and semantics of the text of the second technical document are aligned. The profile can represent a structured summary of the text disclosed in the first technical document. For example, the structured summary can include keywords from the text of the first technical disclosure.

[0045] In some embodiments, a technical document optimization strategy can refer to a set of executable strategy items, including hard constraint strategies and / or soft preference strategies. The soft preference strategy may include one or more of the following items: a strategy for deducing embodiments of the innovative solution, a keyword explanation or overarching term for the innovative solution, a sub-solution decomposition of the innovative solution into a complex solution, a strategy for generating a problem list for the innovative solution, the template and format of the technical document, a strategy for creating diagrams for the innovative solution, a strategy for setting the personal style of the technical document, and a hierarchical (or progressively deeper) description method for the innovative solution from abstract to concrete or from general to specific. Hard constraint strategies may include formatting specifications and standard clauses.

[0046] This application embodiment utilizes a dual-channel strategy engine and language model, uniformly scheduled by a strategy library. Based on optimization strategies, it addresses issues in the first technical document's text, generating a second technical document with a clear structure and semantic alignment. This achieves structural optimization while preserving the original intent of the innovative solution, making the generation process traceable and proactively satisfying strategy constraints. Therefore, the optimized technical document exhibits enhanced readability and understandability, facilitating comprehension and improving the efficiency of subsequent business processing.

[0047] Optionally, as an embodiment, in step 102, the structured summary can be mapped to a decision table for forming the first technical document using the technical document optimization strategy. This decision table includes optimization strategy entries in the technical document optimization strategy that correspond to the structured summary. Alternatively, the profile can be mapped to a decision table for forming the first technical document using the technical document optimization strategy. This decision table includes optimization strategy entries in the technical document optimization strategy that correspond to the profile.

[0048] In one implementation, in step 103, the second technical document can be generated based on the optimization strategy entries included in the decision table.

[0049] Through this scheme, the decision table records the policy items and execution order hit in this case, making the generation process of each second technical document traceable, regressible, and able to be modified in a targeted manner based on user feedback, thereby supporting compliance verification and regression testing.

[0050] In some embodiments, the profile is matched to the hard constraint policy based on a policy engine to obtain a first matching result. In another implementation, the profile is matched to the soft preference policy based on a neural network to obtain a second matching result. The neural network can be a multi-label classifier. The first matching result and the second matching result are fused to generate the decision table. The decision fusion prioritizes the first matching result over the second matching result. Thus, the hard constraint policy must satisfy the requirement that the soft preference policy can operate in regions not covered by the hard constraint policy, and can stably satisfy the hard policy constraints of the technical document.

[0051] In some embodiments, the method may further include: obtaining a third technical document, which is a revised version of the second technical document.

[0052] In one implementation, the uncorrected second technical document is used as the first technical document for the next input to the language model, the output is compared with the third technical document, and the technical document optimization strategy is readjusted based on the comparison result.

[0053] Alternatively, as another embodiment, the method may further include: using the third technical document as the first technical document for the next input of the language model.

[0054] In some embodiments, the second technical document obtained from the i-th revision can be used as the first technical document input to the language model for the (i+1)-th time, where i is a positive integer. Multiple revisions can gradually improve user satisfaction. It should be understood that the embodiments of this application do not limit the feedback triggering conditions, which include, but are not limited to, batch triggering after accumulating N feedbacks, periodic triggering, or manual triggering by the user, where N is an integer greater than or equal to 1.

[0055] Specifically, the feedback iteration loop can asynchronously monitor and collect the user's modification actions on the second technical document. After the user submits confirmation, it triggers the accumulation of the feedback item and the fine-tuning of the language model, so as to avoid frequent fine-tuning of the language model and the strategy library, thereby improving the processing speed.

[0056] In summary, this study can verify whether the updated technical documentation optimization strategy meets expectations. By adjusting the strategy based on the comparison results, the model's capabilities and reliability can be effectively improved, thereby increasing user satisfaction.

[0057] In some embodiments, the difference between the third technical document and the decision table is used as a feedback entry to be precipitated as a new policy entry, expanding the hard constraint policy or the soft preference policy. Specifically, the feedback entry is used as preference alignment data, such as Direct Preference Optimization (DPO) or Reinforcement Learning from Human Feedback (RLHF), to fine-tune the language model, improving optimization performance without retraining a large model.

[0058] According to this implementation method, user modification traces are stored as new strategy entries and preference alignment data, enabling the technical document optimization strategy and the language model to continuously evolve online. This addresses scenarios where the first technical document does not match any existing strategy entries, fundamentally resolving the shortcomings of pure strategy systems (lack of extrapolation) and pure model systems (lack of traceability), ensuring that the strategy library and model capabilities grow together with business accumulation. It should be understood that the embodiments of this application will be illustrated using a technical disclosure document as an example, but can also be any other technical document. Methods or devices using the optimization approach of this application in other technical documents should fall within the protection scope of this application.

[0059] In the following embodiments, the technical document is illustrated using a technical disclosure document as an example. It is not limited to invention patents, but can also be a utility model. The technical document optimization strategy is a disclosure document optimization strategy. The main text of the technical disclosure document includes innovative solutions to the technical problems.

[0060] In some embodiments, the first technical document is a first technical disclosure, and the second technical document is an optimized second technical disclosure. Optionally, the profile can be an invention profile. The main body of the first technical document includes at least one of the following: the type of the innovative solution, the hierarchy of the innovative solution, and the number of entities of the innovative solution.

[0061] Figure 2 This is a flowchart of a technical disclosure document optimization method provided in this application embodiment, which is executed by a technical document optimization device. The method includes the following steps.

[0062] 201: Parse the main text of the first technical disclosure document to obtain a structured summary of the main text of the first technical disclosure document, which includes innovative solutions.

[0063] Optionally, the invention profile can be extracted from the main text of the first technical disclosure document, which includes an innovative solution to the technical problem, and the invention profile includes a structured summary of the innovative solution.

[0064] In some embodiments, an invention profile may refer to a structured summary of the invention disclosed in the first technical disclosure. The structured summary may include keywords from the main body of the first technical disclosure.

[0065] 202: Map the structured summary to a technical disclosure optimization strategy, which includes an optimization strategy corresponding to the structured summary.

[0066] Optionally, the invention profile can be mapped to a technical disclosure optimization strategy, which includes optimization strategies corresponding to the invention profile.

[0067] 203: Generate a second technical disclosure optimization based on the optimization strategy corresponding to the structured abstract. The body of the second technical disclosure optimization includes the original meaning of the innovative solution, and the structure and semantics of the body of the second technical disclosure optimization are aligned.

[0068] Optionally, a second technical disclosure document is generated based on the optimization strategy corresponding to the invention profile, the main text of which includes the innovative solution.

[0069] In some embodiments, the first technical disclosure may be an original technical disclosure provided by the inventor, which at least includes an innovative solution to the technical problem; however, it often has defects in terms of content and / or format. For example, there may be incomplete descriptions of the solution, key parameters being mentioned only briefly, disorganized or unclear narratives, unclear module connections, and / or unclear correspondence between beneficial effects and technical features, typos, and figures that do not conform to patent application specifications, etc., which are areas for improvement.

[0070] In some embodiments, the second technical disclosure document includes the original intent of the innovative solution in the first technical disclosure document, and its text structure and semantic alignment are included. This alignment includes, but is not limited to: key terms with specific explanations, tight contextual connections, fluent sentences, and / or consistent terminology for the same concept throughout the document. Optionally, the text of the second technical disclosure document may also include extended embodiments derived from the innovative solution using neural networks. The text of the second technical disclosure document includes innovative solutions that are essentially the same as those in the first technical disclosure document, but with enhanced readability and comprehensibility or extended features to provide ideas, and its components or chapters are designed to facilitate subsequent business processing.

[0071] An invention profile can include two fields: type and level. The type is a discrete set of labels that answers "what kind of invention this is," such as method / device / apparatus / system / medium, or multi-entity interaction / single-entity, or data-driven / rule-driven, or improved / completely new solution. The level answers "at which level of abstraction is the improvement made," such as link level / system level / network level in communication simulation, module level / service level / platform level in software, device level / circuit level / architecture level in chips, etc. Optionally, the invention profile can also include the number of entities (e.g., single device / single method, interaction between different entities, interaction between N distributed multi-nodes). Furthermore, syntactic dependency analysis and subject clustering can be used to count the number of independent action subjects to initially determine the entity number segment of the invention profile.

[0072] The above-mentioned optimized technical disclosure document scheme adopts a dual-channel approach of "strategy engine and neural network" to process the problems in the main text of the first technical disclosure document according to the optimization strategy and generate a second technical disclosure document. Both technical disclosure documents retain the inventor's original creation, while the optimized technical disclosure document has a structure and semantic alignment, enhanced readability and comprehensibility, facilitates the agent's understanding, reduces communication costs between the agent and the inventor, reduces the deviation of information transmission from the inventor to the agent, and improves the efficiency of business processing.

[0073] In some embodiments, a technical disclosure optimization strategy can refer to a set of executable strategy items related to patent business, including two categories: hard constraint strategies and soft preference strategies. Specifically, hard constraint strategies include compliance with the rigid requirements of patent application law (such as using standardized terms for technical terminology, and avoiding color, photographic, or screenshot-based drawings), unique patent drafting paradigms, standard terminology, and / or standard clauses; soft preference strategies include drafting templates and methods for different types and levels, deductive rules based on the innovative solutions in the original technical disclosure, keyword prioritization, decomposition of complex solutions into sub-solutions, hierarchical progression of solutions from abstract to concrete or from general to specific, drawing creation strategies, question list generation strategies, and / or personal style settings. Hard constraints ensure format compliance, while soft preference strategies achieve patent layout.

[0074] In some embodiments, this hard constraint strategy includes, but is not limited to, formatting specifications and standard clauses. For example, it may specify terminology, prohibit the use of color or photographs in accompanying drawings, and require that drawing references be consistent with the main text of the specification; this hard constraint strategy ensures that the generated results are formatted in compliance with regulations.

[0075] In some embodiments, the soft preference strategy includes one or more of the following: a derivation strategy for embodiments of the innovative solution, a keyword explanation or overarching strategy for the innovative solution, a sub-solution decomposition of the innovative solution into a complex solution, a problem list generation strategy for the innovative solution, a template and format for the technical document, a strategy for creating accompanying drawings for the innovative solution, a strategy for setting the personal style of the technical document, and a hierarchical progressive description of the innovative solution from abstract to concrete or from general to specific. This soft preference strategy can be used to implement patent layout.

[0076] Using the above embodiments, the hard constraint strategy and the soft preference strategy are processed by the strategy engine and the language model, respectively. This avoids problems such as unauthorized rewriting, missing elements, and fabricated data in the pure large model free generation method, so that the generated results can stably meet the constraints of patent drafting.

[0077] For complex technical solutions, the soft preference strategy in the disclosure document optimization strategy, especially the decomposition of complex solutions into sub-solutions, is used. The solution progresses from abstract to concrete or from general to specific levels, which can facilitate reading and understanding, and also facilitate the generation of patent documents for automated systems.

[0078] Inventors often struggle with the realization that their hard work on innovations is wasted if the patent is too narrowly defined, allowing competitors to easily modify and use the technology. A strategy for developing embodiments based on the original technical disclosure's innovative solutions and distinguishing technical features (such as core actions or keywords), prioritizing keywords, and progressively building the solution from abstract to concrete or from general to specific can better define the wording of independent claims and the layout of dependent claims. This avoids narrowing the scope of the inventor's claims and prevents their innovative achievements from being easily circumvented.

[0079] In one embodiment, such as Figure 3 As shown, a decision table can be used to optimize the technical disclosure document. This method includes the following steps.

[0080] 301: Extract the invention profile from the main text of the first technical disclosure document. The main text of the first technical disclosure document includes an innovative solution to the technical problem, and the invention profile includes a structured summary of the innovative solution.

[0081] 302: Map the invention profile to the technical disclosure optimization strategy to form a decision table for the first technical disclosure. The decision table includes the optimization strategy in the technical disclosure optimization strategy that corresponds to the invention profile.

[0082] 303: Generate a second technical disclosure document based on the decision table. The main body of the second technical disclosure document includes the original intent of the innovative solution, and the structure and semantics of the main body of the second technical disclosure document are aligned.

[0083] Furthermore, when the technical document optimization method is applied to the technical disclosure document, it can preserve the inventor's original intention of the innovative solution, and generate a second technical disclosure document with structure and semantic alignment through a dual-channel approach of strategy engine and language model. This effectively reduces the communication costs and information transmission deviations between the agent and the inventor, thereby improving the efficiency of patent drafting.

[0084] Specifically, for example, the mapping logic for querying the optimization strategy corresponding to the invention profile in the technical disclosure optimization strategy and generating the decision table based on the matched optimization strategy is provided; the generation logic for generating the second technical disclosure based on the hard constraint list and soft preference strategy in the decision table is also provided. In some embodiments, the decision table may represent a set of entries for the technical disclosure optimization strategies corresponding to the first technical disclosure, which are matched by the profile (such as the invention profile) in the technical disclosure optimization strategy for the first technical disclosure. The decision table may also include and attach the parameter slots and execution order of each strategy in this case. For example, the decision table may be organized in a JSON structure, and its fields include, but are not limited to: a list of hard constraints (a list of specification terms, a list of figure format constraints, and / or a list of word limits, etc.), an applicable template number (corresponding to the writing template), a recommended extension type (parameterized extension, combined extension, and / or boundary condition extension), and the parameter slots of each strategy in this case. The decision table may also be dynamically updated as the chapter generation progresses.

[0085] The technical solutions provided in this application will now be described in detail with reference to the accompanying drawings. It should be understood that the embodiments of this application are illustrated using a technical disclosure document as an example, but can also be any other technical document. Methods or devices in other technical documents that adopt the optimized approach of this application should fall within the protection scope of this application.

[0086] Figure 4 This is a schematic flowchart illustrating the process of optimizing a technical disclosure method according to an embodiment of this application. The method includes the following steps.

[0087] 401, Obtain the first technical disclosure document and the optimization strategy of the technical disclosure document.

[0088] Optionally, the first technical disclosure document and / or the technical disclosure document optimization strategy may be stored locally or uploaded by the user.

[0089] 402, Analyze the main text of the first technical disclosure document to extract the invention profile.

[0090] In some embodiments, LayoutLMv3 (Layout Language Model version 3) or techniques for recognizing and extracting text content from images, such as OCR (Optical Character Recognition), can be used to parse the main text of the first technical disclosure document, identifying paragraphs, tables, figures, and / or document format conversions. BERT-CRF (Bidirectional Encoder Representations from Transformers-Conditional Random Field) or Span-NER (Span-based Named Entity Recognition) can be used to extract the technical problem to be solved by the innovative solution, the elements of the innovative solution, parameter indicators, and / or figure numbers, and other structured information. A BERT multi-label classifier can also be used to predict type labels and hierarchical labels for key terms in the first technical disclosure document to obtain an invention profile.

[0091] 403, map the invention profile to the optimization strategy of the technical disclosure, and form the decision table of the first technical disclosure.

[0092] The decision table may include optimization strategies (which may be one or more strategy entries corresponding to the invention profile) from the optimization strategy of the technical disclosure. Specifically, it is a mapping logic that queries the optimization strategy corresponding to the invention profile in the optimization strategy of the technical disclosure and generates the decision table from the matched optimization strategy.

[0093] Optionally, a strategy matching approach using a dual-channel parallel approach combining a strategy engine and a neural network can be employed. This involves mapping the invention profile to the technical disclosure document, optimizing the hard constraints and soft preferences of the strategy (which can be achieved using BERT multi-label classification), and then performing decision fusion to generate a decision table.

[0094] Indicative, such as Figure 5As shown, the profile is matched to the hard constraint policy based on the policy engine, resulting in a first matching result; the profile is then matched to the soft preference policy based on a neural network, resulting in a second matching result, where the neural network can be a multi-label classifier. The first and second matching results are then fused to generate a decision table. This decision fusion prioritizes the first matching result over the second matching result. Therefore, this decision fusion can satisfy the hard constraints provided by the policy engine, while the soft inference of the neural network can function in areas not covered by the policy, stably satisfying the hard rule constraints of subsequent document processing (such as patent drafting), and can be executed according to hard constraints, confidence thresholds, or after review by an examiner.

[0095] 404. The second technical disclosure document is generated based on the decision table.

[0096] In some embodiments, a second technical disclosure document can be generated based on one or more strategy entries corresponding to the invention profile in the decision table. The innovative solutions in both the first and second technical disclosure documents are essentially the same, i.e., they retain the inventor's original intent.

[0097] Optionally, the corresponding writing template can be selected according to the applicable template number in the decision table.

[0098] Optionally, the model can generate the components of the second technical disclosure in parallel, chapter by chapter, based on the decision table and the structured elements of the first technical disclosure. Constrained decoding is performed on the output of the language model to satisfy the list of hard constraints in the decision table. Compliance checks are performed on the generated results; if any parts fail, they can be rewritten according to the corresponding strategy entries in the decision table instead of being regenerated entirely, thus preserving the inventor's semantics. The components of the second technical disclosure may include prior art (including, for example, the closest prior art document), technical problems with the prior art, innovative solutions, one or more sub-solutions for implementing the innovative solution, the technical effects of the innovative solution, figures, etc. The second technical disclosure may also include terminology explanations, etc.

[0099] Optionally, based on the implementation strategy of the innovative solution in the first technical disclosure, more implementations can be deduced (which can be based on technical standards or logical reasoning in historical databases, etc.), such as by using simplified RAG (Retrieval-Augmented Generation). Optionally, the extended implementations revolve around the distinguishing features between the innovative solution in the first technical disclosure and the prior art, and do not introduce non-compliant extensions (such as unauthorized rewriting, logical errors, and / or data fabrication, etc.). The extended implementations are not treated as a fact base, but can be used for idea expansion. They can be listed as a problem list so that users (such as inventors or technicians) can later confirm the accuracy of the solution.

[0100] Specifically, distinguishing features can be encoded into query vectors using a query encoder (e.g., dual-tower BERT). Several historical extended instance fragments nearest to the query vector are then retrieved from a vector library (e.g., FAISS or HNSW nearest neighbor search indexes). New extended instances are generated by a language model based on the distinguishing features, nearest neighbor examples, and the extension type (i.e., inference strategy) preference in the decision table, such as through parameterization, combination, and / or boundary condition compliance extensions. Non-compliant extensions can also be filtered out (e.g., content irrelevant to the innovative solution, or a hard constraint list provided by the strategy engine to avoid issues such as unauthorized rewriting, missing elements, and fabricated data under the model's free generation method). These can be listed or marked in the issue list of the second technical disclosure, and can be confirmed or supplemented by the inventor or technical personnel. The issue list may also include corrections made by the second technical disclosure to the first technical disclosure (e.g., missing sentences, incorrect words, alternative expressions of the same technology, and / or typos). Based on the innovative scheme of the original technical disclosure (i.e., the implementation strategy related to the distinguishing technical features), the keyword elevation and the progressive level of the scheme from abstract to concrete or from general to specific, better grasp of the granularity of the elevation can better locate the protection scope of the independent claims and the layout of the dependent claims. In this way, the scope of protection of the inventor's claims can be avoided.

[0101] Optionally, the accompanying drawings provided in the first technical disclosure may be screenshots, photographs, or non-editable images, while the accompanying drawings in the second technical disclosure generated according to the disclosure optimization strategy are editable drawings conforming to the patent application format. In implementation, this embodiment can employ parallel processing of a reserved channel and a redraw channel for the source drawings already included in the first technical disclosure and the newly generated drawings in the second technical disclosure. Specifically, illustrative examples are shown below. Figure 6 As shown, source drawings deemed to be retained are inlined and retained in the second technical disclosure document according to the optimization strategy outlined in the technical disclosure document, such as in vector formats (e.g., Enhanced Metafile (EMF) or Scalable Vector Graphics (SVG)). Text descriptions are generated. Source drawings deemed to be redrawn (e.g., drawings containing trademarks or company logos) are rendered into editable vector graphics using structured code generated by a language model and embedded in the second technical disclosure document. The drawing creation strategy constraints in the disclosure document's optimization strategy are applied to all redrawn drawings. Of course, corresponding drawings can also be generated based on embodiments or extended embodiments of the innovative solution in the first disclosure document (the intuitiveness of the drawings facilitates reading and understanding). These editable graphics can be flowcharts, schematic diagrams, timing diagrams, circuit diagrams, or device structure diagrams for that specific embodiment. In this way, the drawings in the second disclosure document are editable, facilitating subsequent modifications. If the strategy constraints conform to the format requirements of patent drawing, it further facilitates subsequent processing or automatic generation by users, reducing the time cost of drawing creation and improving efficiency.

[0102] 405, Output the second technical disclosure document.

[0103] 406. Obtain the third technical disclosure document, and update the technical disclosure document optimization strategy based on the difference between the third technical disclosure document and the decision table, wherein the third technical disclosure document is the revised second technical disclosure document.

[0104] In some implementations, a revised version of the second technical disclosure document can be obtained from the agent, inventor, or other technical personnel. The difference between this revised version and the decision table is used as feedback entries and deposited as new strategy entries to expand the technical disclosure document optimization strategy (i.e., the hard constraint strategy library or the soft preference strategy library). Furthermore, the language model can be continuously fine-tuned by combining preference alignment methods such as DPO or RLHF, achieving synchronous updates and iterations between the technical disclosure document optimization strategy stored in the memory and the language model. This embodiment uses a feedback loop module to deposit the rewriting traces of the decision table and the second technical disclosure document into new strategy entries and preference alignment data, enabling the technical disclosure document optimization strategy, BERT classifier, and large language model to continuously evolve online without retraining the large model to improve optimization performance. This closed-loop mechanism addresses the scenario where the invention profile of the first technical disclosure document does not match any existing strategy entries, fundamentally solving the two-sided defects of pure strategy systems being unable to extrapolate and pure model systems being untraceable, thereby ensuring that the strategy library and model capabilities grow together with business accumulation.

[0105] Optionally, step 406 can be a feedback iterative loop. For example, the second technical disclosure document revised for the i-th time can be used as the first technical disclosure document as the input for the (i+1)-th time, where i is a positive integer. Through multiple revisions, user satisfaction can be gradually improved.

[0106] Optionally, in one implementation, the subsequent second technical disclosure can be input into the patent automatic generation system after the technology has been confirmed to be correct. It should be noted that the generation of patent application documents can also adopt the innovative solution of the embodiments of this application, using the confirmed second technical disclosure as the first technical document and the patent application document as the second technical document output. The technical document optimization strategy can be formulated according to the characteristics of the patent application document, personal style, or the requirements of different industry fields, thereby improving the efficiency of patent drafting.

[0107] It should be understood that this application does not restrict the specific implementation methods of each step, such as the feedback triggering conditions in step 406 (e.g., batch triggering after accumulating N feedbacks, periodic triggering, and / or manual triggering by the agent, where N is an integer greater than or equal to 1). Optionally, in one implementation, while outputting the second technical disclosure document, asynchronous listening and collection of user (e.g., agent) modification actions on the second technical disclosure document are performed. In another implementation, the trigger can be performed after the user submits confirmation, to avoid frequent fine-tuning of the language model and the strategy library, thereby improving speed.

[0108] In some implementations, the unrevised second technical document (such as the unrevised second technical disclosure document) can be used as the first technical document (such as the first technical disclosure document) input into the system next time. The output document is then compared with the revised second technical document (the revised technical disclosure document), and the technical document optimization strategy is readjusted based on the comparison results (this can be done manually or automatically by the system). This method can verify whether the update of the technical document optimization strategy (such as the disclosure document optimization strategy) meets expectations, and can adjust the technical document optimization strategy based on the comparison results, effectively improving the model's capabilities and reliability, thereby increasing user satisfaction.

[0109] Optionally, in one implementation, the above method may further include updating the disclosure optimization strategy based on the revised second technical document (such as the second technical disclosure document). Optionally, the system can automatically update the disclosure optimization strategy by comparing the results before and after the revision of the second technical document. Of course, the technical document optimization strategy can also be adjusted manually.

[0110] This embodiment can serve not only as a general tool for optimizing disclosure documents across multiple industry sectors, but also as a tool specific to one or more sectors, or unique to an individual, allowing users (not limited to inventors, agents, or technicians, but any user) to implement and iterate on their own. For example, an agent can develop disclosure document optimization strategies based on their accumulated business experience, and then use the solutions or equipment provided in this application to update and iterate them, creating a tool with a personal style, specific to a particular sector, or for a specific company. This provides application flexibility and improves business processing efficiency.

[0111] This solution addresses the difficulty in tracing the decision-making basis for optimization results, thereby enabling seamless integration between technical document optimization and downstream business processes.

[0112] For example, Figure 7 A technical document optimization structure diagram of a language model based on a decoder-integrated self-attention mechanism is shown. It should be understood that... Figure 7The structural diagram is provided to explain the technical solution of this application more clearly and intuitively, and does not limit this application.

[0113] like Figure 7 As shown, the language model sequentially includes a tokenizer, an embedding layer (including token embedding and position embedding), N decoder blocks, an output linear layer, and a normalized output layer (such as a normalized exponential function Softmax layer), where N is an integer greater than or equal to 1. Optionally, N can be greater than 80 and less than 120.

[0114] Each of the N decoding modules comprises a masked multi-head attention layer, a first normalization layer, a feed-forward network (FFN), and a second normalization layer, connected sequentially. Optionally, each decoding module may employ a residual structure; for example, the residual structure may allow the input and output of the masked multi-head attention layer of each decoding module to be used as the input of the first normalization layer, and the input and output of the feed-forward network to be used as the input of the second normalization layer. The feed-forward network layer may use a Rectified Linear Unit (ReLU) or a Gaussian Error Linear Unit (GeLU) as its activation function. Optionally, the first normalization layer may employ methods such as Pre-RMSNorm (Pre-Layer Root Mean Square Normalization) to normalize the input using the root mean square, or it may employ LayerNorm.

[0115] The first technical document and its optimization strategy are input to the word segmentation layer. The word segmentation layer outputs a decision table and the structured elements of the first technical document. In the embedding layer, each token is mapped to a hidden layer dimension, and the absolute position of the token in the sequence is injected into the query / key (Q / K) in the form of a rotation matrix. Long context extrapolation is supported. The output of the embedding layer serves as the input to the masked multi-head self-attention layer of the first of N decoding blocks. After N decoding blocks, the output is sent to the output linear layer (also called the vocabulary projection layer). In this output linear layer, the hidden state is projected onto the vocabulary space. The output of the output linear layer serves as the input to the normalized output layer, which outputs the parallel-generated word sequence of each chapter of the second technical document. Optionally, the output of the normalized output layer can be input to the word segmenter, enabling self-correction and / or sampling diversity.

[0116] The word segmenter can simultaneously receive the first technical document and the technical document optimization strategy, and output a decision table and the structured elements of the first technical document at the word segmenter output. The normalized output layer generates multi-path word sequences in parallel by chapter, and can, for example, decode four chapters in parallel: "existing technology + technical problem / summary of innovative solution / specific implementation / problem list". An additional loop is led out at the output of the normalized output layer to send each output back to the word segmenter input after inverse normalization. This loop allows information between chapters to be cross-referenced during generation, which can reduce terminology drift or possible inconsistencies in content.

[0117] In some embodiments, a constraint decoding module can be connected in series at the output of the language model to ensure that the output satisfies the list of hard constraints in the decision table. Furthermore, guided by the corresponding policy entries in the decision table, sections that fail compliance checks are rewritten in a targeted manner rather than being entirely regenerated, thereby achieving optimization while preserving the semantics of the innovative solution. Through this approach, the language model architecture based on the decoder-integrated self-attention mechanism, in collaboration with constraint decoding, can support the parallel generation of chapters with long context windows and ensure that the generated results consistently satisfy the list of hard constraints.

[0118] Optionally, the technical documentation optimization device can be a server, and the user device can communicate with the server. The technical documentation optimization device can also be embedded in the user device. The user device can be a mobile terminal, such as a mobile phone and a computer with a mobile terminal, for example, a portable, pocket-sized, handheld, computer-embedded, or vehicle-mounted mobile device, including but not limited to mobile devices such as mobile phones or smartphones, personal computers, tablets, iPads, etc.

[0119] In some embodiments, the second technical document includes the original meaning or semantics of the innovative solution in the first technical document (which may be different expressions of the same technology), and its text structure is aligned with its semantics. This alignment of structure and semantics includes, but is not limited to: key terms with specific explanations, tight contextual coherence, fluent sentences, consistent terminology for the same concept throughout the text, and / or the creation of easily understandable diagrams in appropriate locations (such as supplementing a diagram of a solution for greater visual clarity).

[0120] In practice, the inventors trained, tested, and validated the model using several original disclosure documents. For example, they optimized various original disclosure documents, some with defects and others with self-imposed defects. The optimized and original disclosure documents were compared, and the patent application documents generated from the optimized documents were compared with those directly generated from the original disclosure documents. Through several iterations of feedback and trial and error, and by updating the soft and hard strategies in the strategy document, the model's capabilities grew alongside business operations. Related content can be found above and will not be repeated here.

[0121] On the one hand, commonalities among various disclosure documents are sought, such as a constraint strategy that ensures the format or creation of documents that are easy for users to understand and read, and do not deviate from the original inventive intent. On the other hand, the unique characteristics of various disclosure documents are sought, such as strategies driven by the technical content provided in the disclosure documents, given the different technical content of different cases. These strategies include, but are not limited to, dual strategies for two-entity interaction in communication, geometric relationships of mechanical structures, additional components, hierarchical enumeration of higher-level terms, field structure limitations, refinement of core mechanisms, limitations on action implementation, addition of pre- / post- / intermediate steps, and / or triggering conditions. Based on the above solutions, the inventors have also trained and verified several patent application documents, enabling the optimized disclosure documents to interface with patent application documents. This is to ensure that the generated optimized disclosure documents can be integrated into downstream business processes (such as the generation of patent application documents), thereby improving the efficiency of business processing.

[0122] Of course, the solution proposed in this application can also be used to optimize patent application documents. You can refer to the optimization process of the disclosure document mentioned above, which will not be repeated here.

[0123] In this practice, in addition to using existing validation sets for verification, the case of this application was also used for implementation to further verify the generalization ability. The inventors compiled the innovative solution of this application into a technical document as the first technical disclosure. This document was gradually added during the research and development process to implement some technical features of the inventive points. However, it had problems such as disjointed chapters, chaotic layout and lack of context, and the model structure diagram did not meet the requirements of the patent application. The format was also unformatted. The implementation method of this application was used to generate the second technical disclosure. The inventors verified all the contents of the second technical disclosure.

[0124] The second technical disclosure document generated using the scheme of this application includes prior art, technical problems, terminology explanations, innovative solutions (the hierarchical expansion of the technical content can be aligned with the layout of the claims in the patent application and the description of specific embodiments), and drawings. To facilitate technical verification by the inventors, the original text is appended to the appropriate rewritten sections, and corresponding drawings (such as flowcharts) are generated for specific embodiments. The model structure diagram is remade to conform to the drawings required by the patent application. Furthermore, the content is clearly described and logically connected to prevent misunderstandings, and the sentences are fluent and natural. An optimized technical disclosure document can reduce support costs, eliminating the need for significant time spent on the reader's technical understanding and repeated communication.

[0125] The drafting of this application also utilized the scheme outlined in this application. The optimized second technical disclosure was used as the first technical document. A patent application document optimization strategy was developed, and the output second technical document included the invention title, technical field, invention content, specific embodiments, claims, and drawings (i.e., consistent with the format requirements of this application). After review, any discrepancies in the second technical document were addressed with further strategy modifications, and the output second technical document was used as the first technical document for subsequent input. During this process, most of the application content was generated by the system, remaining consistent with the inventor's claimed technical solutions. After review, the patent application document optimization strategy was updated, and the process continued based on the updated strategy (the number of iterations can be determined as needed). Furthermore, it eliminated the need for manually drawing drawings, thereby improving patent drafting efficiency and allowing for continuous experience accumulation.

[0126] It should be understood that the above practice process is only for illustrative purposes to describe the implementation of the solution of this application. The document format, strategy content, and the solution selected in practice (such as whether to iterate or the number of iterations) are all just examples and can be set or selected as needed, or set by individuals, and are not a limitation on the embodiments of this application.

[0127] Figure 8 This is a structural block diagram of a language model-based technical document optimization device provided in an embodiment of this application. The device 800 includes a processor 801 and a memory 802, which are coupled together via a bus system 810. The bus system 810 includes a data bus, and may also include a power bus, a control bus, and a status signal bus. However, for clarity, all buses are labeled as bus system 810 in the figure. Its structure is shown in Figure 8.

[0128] The processor 801 is used to parse the main text of a first technical document (such as a technical disclosure document), extract structured elements, map these structured elements to a technical document optimization strategy stored in memory 802, and generate a second technical document based on the optimization strategy (such as a decision table that can form the first technical document and generate the second technical document based on the decision table). The main text of the first technical document includes a structured summary of the innovative solution. This approach reduces communication costs between the agent and the inventor, minimizes information discrepancies between the inventor and the agent, and improves the efficiency of business processing.

[0129] In some embodiments, for example, the processor 801 can be used to query the optimization strategy corresponding to the structured element in the technical disclosure optimization strategy and generate the decision table from the matched optimization strategy; and to provide generation logic for generating the second technical disclosure based on the list of hard constraints and soft preference strategies in the decision table. In some embodiments, the processor 801 can be used to update the disclosure optimization strategy according to the revised second technical document (such as the second technical disclosure).

[0130] Optionally, the language model-based technical document optimization device 800 may further include a network interface 803, with the processor 801 and memory 802 interconnected via a communication bus 810. Exemplarily, the network interface 803 is used to receive the first technical document, such as from a client, email attachment, or document management system; the network interface 803 is also used to output the second technical document, such as sending it to a user's work platform or a patent application document automatic generation system.

[0131] The technical document optimization device 800 includes, but is not limited to, instances of devices such as servers, workstations, personal computers, cloud computing, and personal terminals that are equipped with natural language processing models and strategy libraries; this application does not limit this to such instances.

[0132] Technical document optimization equipment 800 can achieve execution Figures 1-7 The methods and steps involved are the same as those described in the above method embodiments, and therefore will not be described in detail again to avoid repetition.

[0133] The processor 801 may include, but is not limited to, a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or a central processing unit (CPU), a graphics processing unit (GPU), or a neural processing unit (NPU), or any combination thereof. The memory 802 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory may be random access memory (RAM), used as an external cache. The connection between the processor 801 and the memory 802 can also be implemented via known methods such as shared memory or Network on Chip (NoC), and this application does not limit this.

[0134] In some other embodiments, this application provides a computer-readable storage medium. This computer-readable storage medium stores a computer program that, when executed by one or more processors, implements the method in any possible implementation of the above method embodiments. The computer-readable storage medium may include, but is not limited to, one or any combination of Double Data Rate Synchronous Dynamic Random Access Memory (DDR SDRAM), Static Random Access Memory (SRAM), embedded multimedia card (eMMC), Universal Flash Memory (UFS), Read-Only Memory (ROM), Erasable Programmable ROM, Flash Memory, Solid State Drive, Hard Disk Drive, Optical Disc, Magnetic Tape, Universal Serial Bus Flash Disk (USB Flash Drive), or Secure Digital (SD) Card. The computer-readable storage medium may be non-transitory or transient.

[0135] The above embodiments are used to illustrate the technical solutions of this application and are not intended to limit it. Those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this application without departing from the spirit and scope of this application, and all such modifications and substitutions should be covered within the scope of protection claimed in the claims of this application.

Claims

1. A method for optimizing technical documentation based on language models, characterized in that, The method includes: The structured summary of the first technical document is obtained by parsing the main text of the first technical document, which includes innovative solutions. Mapping the structured summary to a technical document optimization strategy, the technical document optimization strategy including an optimization strategy corresponding to the structured summary; and A second technical document is generated based on the optimization strategy corresponding to the structured summary. The main body of the second technical document includes the original meaning of the innovative solution, and the structure and semantics of the main body of the second technical document are aligned.

2. The method as described in claim 1, characterized in that, The method of mapping the structured summary to the technical document optimization strategy includes: The structured summary is mapped to the technical document optimization strategy to form a decision table for the first technical document, wherein the decision table includes optimization strategy entries in the technical document optimization strategy corresponding to the structured summary; and The step of generating a second technical document based on the optimization strategy corresponding to the structured summary includes: The second technical document is generated based on the optimization strategy entries included in the decision table.

3. The method according to claim 1 or 2, c h a r a c t e r i z e d in that, The technical document optimization strategy includes hard constraint strategy and / or soft preference strategy; in The soft preference strategy includes one or more of the following: the deduction strategy of the embodiment of the innovative solution, the keyword explanation or superposition of the innovative solution, the sub-solution decomposition of the innovative solution into a complex solution, the problem list generation strategy of the innovative solution, the template and format of the technical document, the diagram production strategy of the innovative solution, the personal style setting strategy of the technical document, and the hierarchical progressive description method of the innovative solution from abstract to concrete or from general to specific. The hard constraint strategy includes format specifications and standard clauses.

4. The method of claim 3, wherein, The method further includes: Obtain a third technical document, wherein the third technical document is a revised version of the second technical document; and The differences between the third technical document and the decision table are used as feedback entries to update the technical document strategy by accumulating them into new strategy entries; and The feedback entries are used as preference alignment data to fine-tune the language model.

5. The method of claim 4, wherein, The method further includes: The third technical document is used as the first technical document for the next input into the language model; or The second technical document obtained from the i-th correction is used as the first technical document input into the language model for the (i+1)-th time, where i is a positive integer.

6. The method as described in claim 3, characterized in that, The step of mapping the structured summary to the technical document optimization strategy to form a decision table for the first technical document includes: The structured summary is matched to the hard constraint policy based on the policy engine to obtain the first matching result; The structured summary is matched to the soft preference strategy using a neural network to obtain a second matching result; and The first matching result and the second matching result are fused to generate the decision table; wherein the decision fusion prioritizes the first matching result over the second matching result.

7. The method according to any one of claims 1-6, characterized in that, in The first technical document is a first technical disclosure document, and the second technical document is an optimized second technical disclosure document; wherein the structured summary includes at least one of the following: the type of the innovative solution, the hierarchy of the innovative solution, and the number of entities of the innovative solution.

8. The method as described in claim 7, characterized in that, The step of parsing the body of the first technical document to obtain a structured summary of the body of the first technical document includes: The main text of the first technical disclosure document is analyzed using a layout language model or optical character recognition technology to identify the positions of the main text paragraphs, tables and / or figures; The innovative solution's elements, parameters, and / or graph numbers are extracted using conditional random fields to identify the technical problem it aims to solve; and... The invention profile is obtained by using a multi-label classifier to predict type labels and / or hierarchical labels on the first technical disclosure document.

9. The method as described in claim 7, characterized in that, The step of generating a second technical document based on the decision table includes: Select the corresponding writing template according to the applicable template number in the decision table; Based on the structured elements of the decision table and the main text of the first technical disclosure document, the various components of the second technical disclosure document are generated in parallel by chapter. The various components of the second technical disclosure document include the existing technology and technical problem that is closest to the innovative solution, an overview of the innovative solution, specific embodiments of the innovative solution, and a list of problems. Perform constraint decoding on each component of the second technical disclosure to satisfy the list of hard constraints in the decision table; and Hard constraint verification is performed on each component of the second technical disclosure document, and the parts that fail are rewritten in a targeted manner according to the corresponding strategy entries in the decision table.

10. The method as described in claim 9, characterized in that, The step of generating the components of the second technical disclosure document in parallel by chapter according to the structured elements of the decision table and the main text of the first technical disclosure document includes: The distinguishing features between the innovative solution and the existing technology are determined as the query vector; and Retrieve historical extended embodiment fragments that are nearest neighbors to the queried vector from the vector library; Based on the distinguishing features, the historical extended embodiment fragments and the extended type preferences in the decision table are used to generate extended embodiments according to parameterized extension, combined extension, and / or boundary condition extension; and The extended embodiments are included in the list of issues.

11. The method as described in claim 7, characterized in that, The step of generating the components of the second technical disclosure document in parallel by chapter according to the structured elements of the decision table and the main text of the first technical disclosure document includes: The source drawings in the first technical disclosure document are determined to be retained and are retained inline in the second technical disclosure document according to the vector format in the technical document optimization strategy. If the source drawing in the first technical disclosure document is determined to be a redrawn drawing, an editable vector graphic is obtained by rendering the structured code corresponding to the source drawing and embedded into the second technical disclosure document, and the drawing creation strategy constraints in the technical document optimization strategy are executed; and / or An editable graphic corresponding to a specific embodiment of the second technical disclosure is generated and embedded in the second technical disclosure.

12. The method according to any one of claims 2-11, characterized in that, in The language model is a language model based on a decoder-integrated self-attention mechanism; the language model sequentially includes a word segmenter, an embedding layer, N decoding blocks, an output linear layer and a normalized output layer, where N is an integer greater than or equal to 1; wherein, each decoding block sequentially includes a masked multi-head self-attention sub-layer, a first normalization layer and a second normalization layer. The first technical document is input into the language model, and the second technical document is output based on the language model. Specifically, the word segmenter outputs a decision table and structured elements of the first technical document based on the first technical document and the technical document optimization strategy; through the embedding layer, the N decoding blocks, the output linear layer, and the normalized output layer, the word sequence of each chapter of the second technical document is generated in parallel at the normalized output layer to generate the second technical document, and the output of the normalized output layer is used as the input of the word segmenter.

13. The method as described in claim 12, characterized in that, The value of N is greater than 80 and less than 120.

14. A technical document optimization device based on a language model, characterized in that, The device includes a processor and a memory, the processor, the memory, and the network interface being interconnected via a communication bus; the memory is used to store computer program instructions; the processor is used to read the computer program instructions stored in the memory to perform the method as described in any one of claims 1-13.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, implements the method as described in any one of claims 1-13.

Images