Answer generation method and device based on semiconductor knowledge base and medium

Abstract

The invention relates to the field of semiconductors, in particular to an answer generation method and device based on a semiconductor knowledge base and a medium. The method comprises the steps of performing data association on structured knowledge according to a semiconductor domain knowledge structure to construct a class of knowledge bases; performing data recombination on the unstructured knowledge according to the layout so as to construct a second-class knowledge base; a query question is received, retrieval is executed in the first-class knowledge base and the second-class knowledge base based on a preset retrieval model, and a plurality of related knowledge fragments are obtained; determining a first fusion weight between the knowledge fragments of the same knowledge type based on the relevancy between each knowledge fragment and the query question and the update time of each knowledge fragment; determining an optimal knowledge type according to the group feature information, and determining a second fusion weight between different knowledge types according to the optimal knowledge type; and based on the first fusion weight and the second fusion weight, obtaining a reference answer according to the plurality of knowledge fragments. And the requirements of the semiconductor field on information comprehensiveness and answer accuracy are considered.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of semiconductors, and in particular to a method and device for generating answers based on a semiconductor knowledge base, and a medium. BACKGROUND

[0002] With the rapid development of the semiconductor industry, technology iteration is accelerating, and design complexity is increasing. Enterprises have accumulated a large amount of technical documents, process manuals, fault reports, and research and development notes. However, this knowledge is scattered in different systems, and a semiconductor field knowledge question and answer system is a key requirement for improving research and development efficiency, ensuring production quality, and reducing personnel training costs.

[0003] For example, the patent application with the publication number CN119557279A discloses a database management system and a data management method. The system includes a data server and a first client. The first client is connected to at least one terminal device. The first client is connected to the data server. The data server includes a first database and a second database. The first client is configured to receive the terminal device sent the data to be processed, store and send the data to be processed to the data server. The data server is configured to receive the data to be processed, process the data to be processed of the data type belonging to the text type to obtain the structured data, and store it to the first database. The data to be processed of the data type belonging to the non-text type is processed to obtain the unstructured data, and is stored to the second database.

[0004] For another example, the patent application with the publication number CN117453758A discloses a chip finding method and system based on a knowledge graph and a storage medium. The method includes: crawling chip model, chip original factory, and application field three kinds of entity information to obtain a data source; structuring the unstructured information in the data source, and inserting the structured information into the knowledge graph; constructing a chip knowledge graph database; according to the user's query demand, demand integration, analysis, and translation; the chip knowledge graph database returns the query result to the user. The system includes: a crawler module, a data processing module, a chip knowledge graph, a demand collection module, a finding module, and an output module.

[0005] However, in the application scenario of high precision and high reliability of semiconductors, the accuracy of information directly affects the process stability and production safety. In order to control such risks, the question and answer system usually adopts a conservative building and response strategy, which leads to insufficient scalability of the question and answer system and low practicability. SUMMARY

[0006] The main purpose of the present application is to provide a semiconductor knowledge base-based answer generation method, device and medium. In order to solve the above-mentioned technical problems, the present application specifically adopts the following technical solutions: The first aspect of the present application is to provide a semiconductor knowledge base-based answer generation method, which comprises: S101, obtaining semiconductor knowledge, which is divided into structured knowledge and unstructured knowledge based on knowledge type; S102, based on a first preset model or a preset template, data correlation is performed on the structured knowledge according to general or specific semiconductor field knowledge structure, so as to construct a type of knowledge base; S103, based on a second preset visual model group, data reorganization is performed on the unstructured knowledge according to the layout of the unstructured knowledge, so as to construct a type of knowledge base; S104, receiving a user input query question, performing retrieval in the type of knowledge base and the type of knowledge base based on a preset retrieval model, obtaining several knowledge segments related to the query question; S105, based on the relevance of each knowledge segment and the query question, the update time of each knowledge segment, determining the first fusion weight between each knowledge segment of the same knowledge type; S106, obtaining the group feature information of the user, determining the preferred knowledge type of the user according to the group feature information, and determining the second fusion weight between different knowledge types according to the preferred knowledge type; S107, based on the first fusion weight and the second fusion weight, obtaining a reference answer according to the several knowledge segments.

[0007] The second aspect of the present application is to provide a computer device, which comprises: A memory for storing a computer program; A processor for executing the computer program and implementing the steps of the semiconductor knowledge base-based answer generation method provided by any embodiment of the present application when executing the computer program.

[0008] The third aspect of the present application is also correspondingly provided a computer readable storage medium, the computer readable storage medium stores a computer program, the computer program is executed by the processor, so that the processor implements the steps of the semiconductor knowledge base-based answer generation method provided by any embodiment of the present application.

[0009] Advantages: The embodiment of the application provides a semiconductor knowledge base-based answer generation method, device and medium, the structured and unstructured knowledge base is ensured to have wide breadth by hierarchical construction, and the weight of knowledge segments in the same type and different types is dynamically optimized from multiple dimensions such as relevance, timeliness and user groups, so that the differentiated needs of different users for the accuracy and expansibility of answers are adapted, and the requirements of the semiconductor field for information comprehensiveness and answer accuracy are also considered.

[0010] Firstly, different knowledge base construction paths are adopted for structured and unstructured knowledge, the retrievability and usability of the knowledge are improved, a comprehensive and independent knowledge system is finally constructed by integrating the two types of knowledge, the breadth of the knowledge is ensured from the root, and the coverage and accuracy of the answer are improved. However, the expansion of the breadth will inevitably bring the diversity of the retrieval results, which may contain relevant but non-core or relevant but outdated knowledge segments, and a weight fusion mechanism is introduced to constrain and purify the knowledge segments.

[0011] The first fusion weight in the same knowledge type is determined according to the relevance and update time, and the most relevant and most advanced knowledge is preferentially selected, so that the influence of relevant but low-value or obsolete noise knowledge is effectively reduced. The second fusion weight between different knowledge types is determined based on the group characteristics of the user, the priority of different types of knowledge is adaptively adjusted, and the preference needs of different post engineers or researchers between expansibility and accuracy are adapted. Finally, the process of generating an answer is not simply splicing all the retrieved knowledge segments, but is a weighted fusion under the joint constraint of the two levels of weights, the information completeness advantage brought by the large-scale knowledge base is fully utilized, and the knowledge core is determined through adaptive weight, and the knowledge expansibility and accuracy of the reference answer are balanced. BRIEF DESCRIPTION OF DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, below the drawings needed to be used in the embodiment or prior art description will be simply introduced. In all the drawings, similar elements or parts are generally identified by similar reference signs. In the drawings, each element or part is not necessarily drawn according to the actual proportion. Obviously, the drawings described below are some embodiments of the application, and those skilled in the art can obtain other drawings according to these drawings without paying creative labor.

[0013] Figure 1 is a schematic flow chart of a semiconductor knowledge base-based answer generation method provided by the embodiment of the application; Figure 2 is a constituting schematic diagram of a second preset visual model group provided by the embodiment of the application; Figure 3is a schematic diagram of a semiconductor knowledge base provided by an embodiment of the present application; Figure 4 is a schematic flow chart of an answer output method based on a semiconductor knowledge base provided by an embodiment of the present application; Figure 5 is a schematic flow chart of an answer output mechanism under risk control provided by an embodiment of the present application; Figure 6 is a schematic flow chart of a building method of a semiconductor knowledge base provided by an embodiment of the present application; Figure 7 is a schematic flow chart of a retrieval method of a semiconductor knowledge base provided by an embodiment of the present application; Figure 8 is a schematic block diagram of a structure of a computer device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0014] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in detail with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0015] The flow charts shown in the drawings are only exemplary and do not necessarily include all the contents and operations / steps, nor are they necessarily executed in the order described. For example, some operations / steps can be further divided, combined or partially merged, so the actual execution order can be changed according to the actual situation.

[0016] In this document, the suffixes such as "module", "part" or "unit" used to represent elements are only for the convenience of description of the present application and do not have specific meanings. Therefore, "module", "part" or "unit" can be used interchangeably.

[0017] In this document, the terms "upper", "lower", "inner", "outer", "front", "back", "one end", "the other end" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of description of the present application and simplification of the description, and do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second" are only for the purpose of description and cannot be understood as indicating or implying relative importance.

[0018] In this document, unless otherwise indicated and limited, the terms "mount", "provided with", "connected", and the like, should be interpreted broadly, for example, "connected" can be fixed connection, can also be detachable connection, or integral connection; can be mechanical connection, can be direct connection, can also be indirect connection through intermediate medium, can be internal communication of two elements. For those skilled in the art, the specific meaning of the above terms in this application can be understood according to the specific circumstances.

[0019] In this document, the term "and / or" includes any and all combinations of one or more listed associated items.

[0020] In this document, the term "a plurality of" means two or more, that is, it includes two, three, four, five, etc.

[0021] It should be noted that in this document, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, method, article or device. Without more limitations, the element defined by the sentence "including a…" does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0022] With the rapid development of the semiconductor industry, the technology iteration cycle is continuously shortened, and the complexity of chip design and manufacturing is continuously improved. Enterprises accumulate a large amount of technical documents, process specifications, fault reports, and research and development notes and other materials in the research and development, production and operation processes. These materials are usually scattered in different information systems or departments, lack of unified management and effective integration, resulting in low knowledge reuse efficiency and high information search cost. Under this background, building a professional question and answer system for the semiconductor field has become an important demand to improve research and development efficiency, ensure production stability, and reduce personnel training cost.

[0023] In order to enhance the knowledge coverage ability and scalability of the question and answer system, it is usually necessary to integrate the knowledge content into the knowledge base of the question and answer system. However, there are obvious contradictions in the knowledge integration process: if the knowledge breadth is excessively pursued, a large number of documents are included, which may easily lead to a large amount of low-relevance or redundant information in the search results, and then affect the accuracy and conciseness of the answer generation; on the contrary, if the search range or knowledge access standard is set too strictly, some technical details that are not core but have actual reference value may be excluded, causing information omission and leading to one-sided answer generation.

[0024] In scenarios such as semiconductor research and manufacturing, which have extremely high requirements for precision and stability, information accuracy is directly related to production safety and product quality. For example, if the question and answer system provides incorrect process parameter recommendations, it may cause abnormal device performance or yield reduction; misjudgment of equipment failure may cause unplanned downtime, and in severe cases, it may even affect the normal operation of the entire production line, causing significant economic losses. Therefore, to control risks, existing question and answer systems generally adopt conservative strategies in knowledge storage and answer generation, only opening response permissions for content that has been fully verified, and limiting the output of uncertain information, which leads to a longer knowledge update cycle, limited response capability, and limited coverage scenarios in the semiconductor field, making it difficult to meet the query needs of engineers.

[0025] Based on this, the embodiments of the present application provide a semiconductor knowledge base-based answer generation method, device and medium, which ensures the breadth of the knowledge base by constructing structured and unstructured knowledge bases in layers, and dynamically optimizes the weights of knowledge fragments within the same type and between different types from multiple dimensions such as relevance, timeliness and user groups, to adapt to the differentiated needs of different users for the accuracy and expansiveness of answers, while taking into account the requirements of the semiconductor field for information comprehensiveness and answer accuracy.

[0026] The embodiments of the present application also provide a semiconductor knowledge base-based answer output method, device and medium, specifically proposing a differentiated answer output mechanism under strict risk control, which improves the safety and practicality of the question and answer system in the high-precision semiconductor field, effectively avoids risks (such as line downtime) caused by incorrect information, and maintains the response efficiency of the question and answer system.

[0027] Some embodiments of the present application will be described in detail below with reference to the accompanying drawings. In the case of no conflict, the embodiments described below and the features in the embodiments can be combined with each other.

[0028] Please refer to Figure 1 , Figure 1 is a schematic flowchart of a semiconductor knowledge base-based answer generation method provided by the embodiments of the present application, as Figure 1 indicated, the embodiments of the present application provide a semiconductor knowledge base-based answer generation method.

[0029] S101, obtain semiconductor knowledge, which is divided into structured knowledge and unstructured knowledge based on knowledge type.

[0030] Specifically, the semiconductor knowledge to be stored is classified according to its inherent organizational characteristics, divided into two basic categories of structured type and unstructured type, thereby obtaining structured knowledge and unstructured knowledge. Among them, the structured knowledge refers to the content with clear data mode and fixed field format, such as process parameter table, etc.; the unstructured knowledge refers to the content without fixed format, such as technical documents, fault analysis reports, research and development notes, etc. By identifying the structure type to which the knowledge belongs, the corresponding analysis method and subsequent storage management strategy can be targeted, thereby improving the efficiency and accuracy of knowledge processing.

[0031] In specific application scenarios in the semiconductor field, structured knowledge mainly manifests as data with fixed mode, clearly defined fields and relationships, and its carriers include but are not limited to: table files in CSV format, hierarchical data files in JSON format, or relational / non-relational database records complying with specific architecture. Unstructured knowledge widely exists in documents lacking fixed format or organization, and its carriers include but are not limited to: PDF documents, WORD documents, TXT plain text files, or Markdown files containing semi-structured text.

[0032] In some embodiments, a knowledge base space can be created, and the construction of an initial knowledge base space can be realized through a specific function module, in which a user or system administrator defines and sets the core attributes and parameters of the knowledge base. These parameters at least include: a unique name identifier given to the knowledge base, access and use permission rules (such as public, private or specific user group) specified for the knowledge base, information of the main creator of the knowledge base, selection of a large language model (LLM) relied on by the knowledge base for processing text data, determination of an Embedding model for generating knowledge vector representation, and specific model types or service sources of visual detection and multi-modal models for document parsing. The setting of these basic information is the starting point for the establishment and operation of the knowledge base system, and provides necessary framework and resource configuration for subsequent knowledge processing. Among them, LLM is a natural language processing system based on deep learning, which can capture complex semantic structures, long-distance dependencies and world knowledge of human language through self-supervised learning. Embedding model is an AI technology that maps discrete data (such as words, sentences or images) to continuous vector space, and in natural language processing, the most common form is text Embedding, which converts text into high-dimensional vectors. For details, please refer to related technologies.

[0033] In S102, based on a first preset model or a preset template, the structured knowledge is data-associated according to general or specific semiconductor field knowledge structure, to construct a type of knowledge base.

[0034] The general or specific semiconductor field knowledge structure refers to a knowledge organization framework that has been agreed upon and fixed for use within the semiconductor industry or enterprise, including clear entity types (such as devices, processes, materials, and equipment), relationship types between entities (such as using materials and using processes), and attributes associated with various entities and relationships (such as device size parameters and process temperature ranges). The semiconductor field knowledge structure can be flexibly defined and adjusted based on industry standards or internal technical specifications, and is not limited here.

[0035] The first preset model is a model for extracting the association between different parameters in structured knowledge, which is suitable for data sources with relatively strong generality and standard data structure. This model is optimized or trained according to the general or specific semiconductor field knowledge structure.

[0036] By way of example, the first preset model can use LLM. In LLM, prompt engineering (also known as contextual prompt) is a method of guiding LLM to generate a specific type of response by not updating the weights or parameters of the model. Prompt refers to the input instructions, questions, or text prompts. By introducing actual business knowledge in the semiconductor field (such as general or specific semiconductor field knowledge structure), the prompt is designed to clearly specify the entity types, relationship types, and related attributes, and to standardize the output format in the prompt. On this basis, LLM can analyze structured data according to the general or specific semiconductor field knowledge structure, automatically identify and extract the triples composed of the subject, relationship, and object, and output in the required format.

[0037] The preset template is a standardized organization format designed based on the general or specific semiconductor field knowledge structure, providing a standardized field matching mode. The same type of structured data is uniformly aligned and mapped when input, and is suitable for specific structured data with clear hierarchical relationships, fixed structure, and high reusability, such as the basic information and hierarchical relationships of databases, tables, and fields in a relational database. A preset template can be set.

[0038] By way of example, the user can manually define the required entity types (such as "database", "table"), relationship types (such as "contains table", "owns field"), and attribute information of various categories (such as database name, table name, field name, field type) through a configuration interface. These rules and mapping logic are hard-coded (i.e., fixed program logic is written) and fixed, forming a customizable template file that can be called repeatedly, i.e., a preset template. When processing the same type of structured data later, the hard-coded rules in the existing preset template can be used to automatically parse the input structured data, quickly building a knowledge graph containing entities, relationships, and basic attributes, without the need for repeated configuration or reliance on LMM reasoning.

[0039] It should be understood that the structured knowledge carrier itself has clear data item division, clear relationship definition, data type and attribute constraint, and other standardized features. It is difficult to support the deep relationship reasoning requirements that may be involved only by relying on the original data structure. In order to improve its retrievability and reasoning ability in complex scenarios, the mechanism of combining LLM and customized templates is adopted to perform deep relationship analysis before knowledge storage, in order to solve the shortcomings of vector retrieval and keyword retrieval in processing long-distance dependence and implicit association, such as the indirect influence of process steps on device performance.

[0040] In some embodiments, the S102 further includes: if the structured knowledge is a preset specification data, calling a preset template corresponding to the specification data to perform content analysis on the structured knowledge to generate a type of data corresponding to the structured knowledge; if the structured knowledge is not a preset specification data, calling the first preset model to perform content analysis on the structured knowledge to generate a type of data corresponding to the structured knowledge.

[0041] Among them, the preset specification data refers to the structured data conforming to the standard format defined by the enterprise or industry, and the field naming, hierarchical relationship and data type thereof follow fixed rules, such as standardized process parameter table or device specification list, enterprise internally defined list format, which can be flexibly customized according to actual needs, and is not limited here.

[0042] Specifically, if the input data belongs to the preset specification data, there is a customized template file, and the matched preset template is directly called for analysis, and the entities and relationships are automatically extracted through the pre-hard-coded template to generate a type of data; if the structured knowledge format is not fixed or no template is defined, the first preset model is called, and the understanding ability trained based on the knowledge structure in the semiconductor field is used to guide the large language model to identify the triplets of entities, relationships and objects that may exist, and output a type of data that meets the requirements. Therefore, the processing efficiency and adaptability are considered, and it is ensured that structured knowledge of different sources can be effectively converted into a type of data with high retrievability.

[0043] S103, based on the second preset visual model group, reorganizing the unstructured knowledge according to the layout of the unstructured knowledge to construct a second type of knowledge base.

[0044] The second preset visual model group refers to a group of trained models for analyzing the visual layout of a document, which can identify the spatial distribution and content information of elements such as text, tables, and images in unstructured knowledge. For example, typical technical document samples in an enterprise are collected, such as process instructions and fault reports in PDF format, which contain complex and diverse heterogeneous elements such as natural text, images, tables, mathematical formulas, headings, and page header / footer content. Image processing techniques are used to convert the samples into visual inputs for analysis and data annotation to form a training data set. The second preset visual model group is trained on these training data sets to learn the spatial distribution and content extraction of various heterogeneous elements.

[0045] Specifically, for unstructured knowledge, to effectively extract complete and high-quality information contained therein, the spatial layout relationship and type differences between various heterogeneous elements must be fully considered and processed, and the second preset visual model is used for targeted analysis. First, existing analysis tools (such as PDF analysis tools) are used to convert the page content of unstructured knowledge into high-resolution image sequences, and the second preset visual model group is used to identify the boundaries and types of various heterogeneous elements. On this basis, OCR technology is used to extract the content in each region, such as parsing the content of table elements into corresponding fields and numerical values. Then, the extracted content information is reorganized according to the actual layout to form two-class data that can be used for subsequent storage and retrieval, thereby constructing a two-class knowledge base.

[0046] In some embodiments, the second preset visual model group includes a layout detection model and a multi-modal recognition model; S103 includes: the second preset visual model group includes a layout detection model and a multi-modal recognition model; S103 includes: based on the layout detection model, performing layout analysis on the unstructured knowledge to obtain a plurality of blocks and block layout data; based on the multi-modal recognition model, performing content analysis on blocks of different modalities respectively to obtain block content data of each block; based on the block layout data, splicing and reorganizing a plurality of block content data to obtain two-class data corresponding to the unstructured knowledge.

[0047] Exemplarily, the unstructured knowledge document is converted into an image format and input into a layout detection model, the layout detection model performs layout element analysis on the images, accurately detects the area position, boundary and hierarchical relationship of text blocks, image blocks, table blocks, formula blocks and element regions such as titles, headers or footers, and forms block layout data. The multi-modal recognition model refers to a model capable of processing multiple data modalities (such as text state, formula state, table state) and extracting the content thereof. The multi-modal recognition model is called to perform classification processing on each block (such as a text paragraph, a parameter table and a process schematic diagram) respectively, and obtain the content data of each block. According to the spatial and logical order in the block layout data, the block content data of each block is spliced and integrated according to the context relationship, to generate coherent and clearly structured two-class data.

[0048] In some embodiments, the block content data includes text data in a text form, formula data and table data; the multi-modal recognition model includes: a text recognition model configured to recognize and analyze block knowledge existing in a text state to obtain the text data; a formula recognition model configured to recognize and analyze block knowledge existing in a formula state to obtain the formula data; and a table recognition model configured to recognize and analyze block knowledge existing in a table state to obtain the table data.

[0049] Specifically, for the detected table and formula elements, the table data is finally converted into a structured text expression form (such as a Markdown format), the formula data is converted into an accurate LaTeX format text, and the remaining non-table / formula text elements (such as a normal text paragraph, a title, a header and a footer) are recognized by the text recognition model (such as an OCR model) and output as pure text.

[0050] In some embodiments, the block content data is spliced based on the block layout data, including: according to the coordinate position information (such as a top-left corner coordinate) of all analyzed elements in the original page, the block content data (such as text from text OCR, Markdown text after table analysis and LaTeX text after formula analysis) is sorted and spliced in a preset order (such as a physical reading logical order from left to right and from top to bottom), to finally generate a high-quality structured text output integrating all visual elements.

[0051] Please refer to Figure 2 , Figure 2 is a constituting schematic diagram of a second preset visual model group provided by the embodiment of the present application, as shown in Figure 2As shown, the second preset visual model group includes: a layout detection model responsible for accurately identifying the position bounding box and category of various elements in the page; a multi-modal recognition model group according to the block type, which can include: a formula recognition model responsible for recognizing and parsing mathematical formulas; a table recognition model responsible for recognizing and parsing table structures; and a text recognition model (such as an OCR model) responsible for recognizing the text in the image area. It should be understood that the embodiments of the present application adopt a parsing mechanism that combines two types of models, visual detection and multi-modal recognition. The layout detection model and the text recognition model belong to visual detection according to their recognition mechanism, which realizes accurate separation and structured reorganization of multiple types of information in complex layouts, and improves the integrity and accuracy of unstructured knowledge parsing.

[0052] In order to meet the subsequent visualization inspection requirements and the needs of subsequent search enhancement stage which may integrate different search strategies (such as vector search, keyword full-text search, and graph relationship query), a multi-modal storage scheme is provided. Please refer to Figure 3 , Figure 3 is a schematic diagram of a semiconductor knowledge base provided by an embodiment of the present application, which includes a semiconductor knowledge base and a second type of knowledge base.

[0053] In some embodiments, the first type of knowledge base includes a knowledge graph database, and the first type of data is knowledge graph data; the second type of knowledge base includes a vector database and / or a full-text database; and the method includes: storing the second type of data in the vector database after vectorization processing, and / or storing the second type of data in the full-text database.

[0054] Specifically, for the first type of data obtained by structured knowledge parsing, a knowledge graph database (also known as a graph database) is used for storage, which supports efficient structured relationship and path query; for the second type of data obtained by unstructured knowledge parsing, the text form of various block content data is vectorized to obtain vectorized representation and stored in a vector database to support vector search based on semantic similarity; and a full-text database is used to store original text or processed text and establish an inverted index to support accurate keyword or Boolean logic matching search.

[0055] In some embodiments, the original document, intermediate or final visualization inspection required files (such as processed PDF images, layout analysis result images) generated during the parsing process can be stored in a file object storage service for subsequent synchronization output as search results with reference answers, which facilitates intuitive comparison and verification by users.

[0056] In some embodiments, a visual inspection mechanism for human-machine interaction is provided to ensure the accuracy and integrity of the parsing process. The user can enter a designated knowledge base space and randomly select any document that has been stored. The system will extract the original document (or its high-fidelity presentation form) from the file object storage service and display it side by side with a class of data or two-class data on the user interface, such as the parsed, structured and fused sorted high-quality text content or key elements (such as the parsing results of tables and formulas), to realize intuitive comparison of the original content and the parsing results. The user can quickly and clearly identify and judge whether there are quality problems such as missing content information, text recognition errors, table or formula conversion errors, layout logic disorders (such as element order reversal) in the parsing process by directly visually comparing the two content presentations.

[0057] In some embodiments, a closed-loop correction mechanism based on visual inspection results is provided. When the user finds problems in the parsing results through visual comparison, they have the right to directly modify or supplement the incorrect content through designated channels. The corrected data record will be marked with its source (such as the modifier ID) and time. Subsequently, the corrected version is submitted to a superior manager or a designated domain expert (such as a semiconductor process expert or a data expert) for review and evaluation. The reviewer checks the correctness and compliance of the correction content. If the review conclusion is passed, the data of the corrected version will be officially updated to the corresponding knowledge base storage system (including the vector library, full-text retrieval library, graph database or file storage content that may be involved), covering the original incorrect parsing results; if the review conclusion is not passed, the system will feedback the specific opinions of the review expert to the original modifier. The modifier needs to modify the problem again according to the expert's opinions to form a new round of correction version and submit it to the review process again. This "correction-review-update" or "correction-feedback-re-correction-review-update" cycle will continue until the review is passed, ensuring that the final stored knowledge has high accuracy and authority.

[0058] S104, receiving a query question input by a user, performing retrieval in the first knowledge base and the second knowledge base based on a preset retrieval model, and obtaining a plurality of knowledge segments related to the query question.

[0059] Among them, the preset retrieval model refers to an algorithm or strategy combination configured in advance according to application requirements for performing queries in different types of knowledge bases. It can call the appropriate retrieval method for the knowledge graph database, vector database and full-text database to obtain knowledge segments related to the user's query question.

[0060] Specifically, a natural language query question input by a user is received, the question is standardized and preprocessed, including text cleaning and semantic normalization, a first type of knowledge base and a second type of knowledge base are accessed in parallel based on a preset retrieval model. In the knowledge graph database, relevant triple structures are found through entity recognition and relationship matching mechanisms; in the vector database, the problem is vectorized and approximate nearest neighbor search is performed to obtain knowledge fragments with similar semantics; in the full-text database, keyword matching and Boolean logic retrieval are performed using an inverted index. Finally, the retrieval results from the multiple source knowledge bases are summarized, sorted by relevance, and a set of knowledge fragments associated with the query question is formed to provide input for subsequent answer generation.

[0061] S105, based on the relevance of each knowledge fragment to the query question and the update time of each knowledge fragment, determine the first fusion weight between each knowledge fragment of the same knowledge type.

[0062] Specifically, after completing retrieval from a first type of knowledge base (such as a knowledge graph database) or a second type of knowledge base (such as a vector database or a full-text database), the first fusion weight within each knowledge fragment set returned by each type of knowledge base is calculated independently. The first fusion weight refers to a coefficient used to measure the contribution proportion of each knowledge fragment in subsequent answer generation within the same knowledge type, which is used to sort and weight the same type of knowledge fragments retrieved from the first type of knowledge base or the second type of knowledge base. The higher the first fusion weight, the higher the contribution proportion of the knowledge fragment, which improves the accuracy and timeliness of the generated reference answer.

[0063] In some embodiments, the first fusion weight can be calculated based on semantic relevance, content timeliness, source reliability, and other evaluable factors, and can be a weighted sum of each evaluation dimension. For example, the relevance of each knowledge fragment to the query question is the semantic relevance, which is output by the preset retrieval model when performing retrieval, such as vector similarity score, keyword matching strength, or graph path matching depth. The semantic relevance score can be normalized to the [0, 1] interval, and a higher score indicates a closer semantic association. For example, the update time is used to evaluate the content timeliness of the knowledge, the creation or last revision time of each knowledge fragment is obtained, and the timeliness score of the knowledge fragment is calculated according to the current time. The more recent the update time, the higher the timeliness score, for example, knowledge within one year has a score of 0.9 or higher, and gradually decreases over time, such as knowledge more than three years old decreases to 0.6, so that newly recorded process specifications or technical reports have higher weights. For example, source reliability is introduced as a supplementary dimension, for example, knowledge entries from enterprise standard documents or expert-reviewed knowledge entries can be assigned a higher source score, while content from personal notes or unverified reports is appropriately reduced.

[0064] The overall score is obtained by linearly weighting the scores of all dimensions, such as overall score = a x semantic correlation score + b x timeliness score + g x source score; wherein a, b, and g are adjustable weight coefficients, and the sum is 1, which can be flexibly configured according to the application scenario. For example, in the process parameter query scenario, the weight of the timeliness score can be increased, and a = 0.5, b = 0.4, and g = 0.1 are set. In the historical fault analysis, the weight of the source score can be appropriately increased, and a = 0.5, b = 0.2, and g = 0.3 are set. Finally, the overall scores of the various knowledge segments are normalized (i.e., the overall score of the knowledge segment is divided by the sum of the overall scores of the same type of knowledge segment), and the result is the first fusion weight of each segment.

[0065] In S106, the group characteristic information of the user is obtained, the preferred knowledge type of the user is determined according to the group characteristic information, and the second fusion weight between different knowledge types is determined according to the preferred knowledge type.

[0066] The second fusion weight refers to a coefficient for measuring the contribution proportion of each knowledge segment in subsequent answer generation between different knowledge types (i.e., between the structured knowledge of a type of knowledge base and the unstructured knowledge of a type of knowledge base), which is determined based on the preferred knowledge type of the group to which the user belongs and a preset rule mapping, and the value range is [0, 1], and the sum of the weights of the two types of knowledge is 1.

[0067] Specifically, the preferred knowledge type of the user is determined according to the group characteristic information of the user, and then the second fusion weight is set. For example, for a process engineer, who is responsible for formulating and optimizing the chip production process flow, mainly involves parameter debugging and verification of key processes such as lithography, etching, deposition, and ion implantation, and highly depends on accurate and executable parameter instructions in daily work, therefore, more trust is given to structured data such as process specifications and equipment configuration tables, and such users are marked as preferring structured knowledge, and the second fusion weight is set as: the weight of structured knowledge is 0.8, and the weight of unstructured knowledge is 0.2, so that in generating answers, the process parameter triples from the knowledge graph database will occupy a dominant position, and the unstructured data from the fault report or the research and development notes will only serve as supplementary explanation. For example, for yield analysis engineers or product development engineers, their work often needs to refer to historical cases, abnormal processing records, and expert experience, and therefore, they rely more on unstructured knowledge, and are marked as "prefer unstructured knowledge", and the second fusion weight is set as: the weight of structured knowledge is 0.4, and the weight of unstructured knowledge is 0.6. For example, a device operation and maintenance engineer may need to consider both theoretical and measured data, and the balanced weight can be set as: the weight of structured knowledge is 0.5, and the weight of unstructured knowledge is 0.5.

[0068] In some embodiments, the group feature information refers to the post type (such as process engineer, equipment engineer, R&D engineer) of the user in the semiconductor enterprise, the department (such as front-end process, lithography, integration, yield analysis) to which the user belongs, the authority level, and the like. For example, the identity information of the user can be automatically synchronized from the internal system of the enterprise to obtain the basic attributes such as the post, department, and job level of the user. In addition, the historical operation behavior of the user can be analyzed, for example, the high-frequency query content of the user can be analyzed to further confirm the relevant post of the user, for example, the user frequently searches for fault codes and maintenance records, and it is more likely that the user is in a post related to equipment maintenance.

[0069] In some embodiments, a group and preference mapping rule library is preset, and the preferred knowledge type of the user can be quickly obtained according to the identified group feature information.

[0070] In some embodiments, the S106 further includes: obtaining an initial fusion weight, wherein the weight of the structured knowledge segment in the initial fusion weight is higher than the weight of the unstructured knowledge segment; and adjusting the weight of the preferred knowledge type in the initial fusion weight to update the initial fusion weight to obtain the second fusion weight.

[0071] In some embodiments, the initial fusion weight refers to a default fusion ratio preset based on the inherent reliability of the knowledge type before considering the user preference. The structured knowledge can be given a higher basic weight, for example, the first type of knowledge is 0.6 and the second type of knowledge is 0.4.

[0072] It should be understood that the setting of the initial fusion weight is based on the essential difference of the two types of knowledge in data source, processing link and reliability. The first type of knowledge is derived from structured data such as process parameter table, device configuration library, device specification database, etc. The original data has the characteristics of standard format, clear field and accurate value, and usually comes from verified enterprise standard documents or industry general standards, with high accuracy and consistency. In the construction process of the first type of knowledge base, a preset template is preferentially used for parsing, and entities, attributes and relationships are directly extracted through hard-coded rules. The processing link is data correlation, that is, semantic mapping is established on the basis of existing structure. This process has low dependence on large language models, strictly follows the original data, has less algorithm intervention and small error risk, so the knowledge graph data formed finally has strong credibility and stability. In contrast, the data of the second type of knowledge base comes from unstructured data such as technical manuals, research notes, academic papers and fault analysis reports. Its content is free-form and has higher knowledge depth, and may have ambiguous expressions, unverified conclusions or rely on specific scene conditions. Its processing link is data reorganization, which needs to use the layout detection model and multi-modal recognition model in the second preset visual model group to complete layout analysis, semantic extraction and reconstruction, involving more processing links and model reasoning, and having higher possibility of misidentification or semantic deviation. Therefore, although unstructured knowledge has high value in supplementing background and experience summary, its overall reliability is still lower than that of structured knowledge.

[0073] Based on this, in the initial stage, the structured knowledge in the first type of knowledge base is preferentially trusted, and based on this, the second fusion weight is obtained by increasing the weight of the preferred knowledge type determined according to the user group characteristic information. The increase ratio can be configured according to the actual use feedback of the enterprise, and can be adjusted within the range of ±0.1 to ±0.3, which can adapt to the information preferences of different groups while maintaining overall stability. It should be understood that in general, the weight of structured knowledge in the adjusted second fusion weight is still dominant, and the weight of unstructured knowledge may be higher in specific scenarios, thereby balancing the preference of different users for knowledge expansion while ensuring core accuracy.

[0074] For example, if the user is a process engineer or other group that relies on accurate parameters, the weight of structured knowledge is further increased from 0.6 to 0.7, and the weight of unstructured knowledge is reduced to 0.3, forming the final second fusion weight. For example, if the user is a research and development or analysis personnel who needs to refer to experience-based information, the fusion proportion of unstructured knowledge is moderately increased from 0.4 to 0.5, and the weight of structured knowledge is correspondingly reduced to 0.5, forming the final second fusion weight, which balances the user's individual needs while ensuring the reliability of the answer.

[0075] S107, obtaining a reference answer according to the knowledge segments based on the first fusion weight and the second fusion weight.

[0076] Specifically, when generating an answer based on each knowledge segment using a large language model, the knowledge segment with a higher first fusion weight and second fusion weight is given priority. This can be achieved by adjusting the attention mechanism or prompt of the model, for example, setting the prompt as "generate the answer based on structured knowledge first, and unstructured knowledge is only for reference".

[0077] For example, the query question is "What is the standard material for device M?". From a type of knowledge base, knowledge segments are retrieved: segment A (such as the 2024 process specification) has a first fusion weight of 0.85, and the relevant content extracted is "the standard material for device M is chemical substance X"; segment B (such as the 2022 fault report) has a first fusion weight of 0.70, and the relevant content extracted is "there is a certain probability that the use of material X in device M will cause device defects". From a second type of knowledge base, knowledge segments are retrieved: segment C (such as the 2023 research and development notes) has a first fusion weight of 0.80, and the relevant content extracted is "device M uses material Y to reduce costs"; segment D (such as the 2023 process summary report) has a first fusion weight of 0.75, and the relevant content extracted is "manufacturing of device M should pay attention to thermal stress dimension".

[0078] If the user group is a process engineer, set the second fusion weight as 0.7 for structured knowledge and 0.3 for unstructured knowledge. At this time, the first fusion weight and the second fusion weight of segment A of the large language model are higher, and "chemical substance X" is the current standard answer, while limited reference to unstructured data is only mentioned under the premise of not interfering with the main conclusion. For example, the following reference answer is generated: "The commonly used material for device M is chemical substance X, and early production processes may cause defects, so the manufacturing process needs to pay attention to thermal stress dimension". In this way, while ensuring accuracy and authority, relevant technical reference information is retained.

[0079] If the user group is a research and development engineer, set the second fusion weight as 0.4 for structured knowledge and 0.6 for unstructured knowledge. At this time, the following reference answer is generated: "Device M mainly uses chemical substance X as the material, but historical data shows that there may be a risk of defects; and material Y has a cost advantage and can be further evaluated." In this way, while ensuring accuracy and authority, potential technical reference information is retained.

[0080] In some embodiments, for the set of knowledge segments retrieved from the first type of knowledge base and the second type of knowledge base respectively, when the number of retrieved knowledge segments is greater than a preset number, the first fusion weight of each is applied for sorting and filtering, and the top preset number of knowledge segments with the highest weight in each type of knowledge are retained.

[0081] In some embodiments, the S107 further comprises: based on the first fusion weight, integrating and splicing knowledge pieces of the same knowledge type into a target knowledge piece; based on the second fusion weight, weighting and fusing the target knowledge piece to generate the reference answer.

[0082] Specifically, based on the first fusion weight, the knowledge pieces of the same knowledge type are internally integrated, and the multiple structured knowledge pieces from a type of knowledge base are sorted according to the first fusion weight. The pieces with a weight higher than a threshold value are selected and spliced into a unified target knowledge piece in a logical order. For example, the related contents of piece A and piece B are integrated into a first target knowledge piece: "The commonly used material of device M is chemical substance X, and the early production process may cause defects", and the related contents of piece C and piece D are integrated into a second target knowledge piece: "The material of device M can try material Y, and the manufacturing process should pay attention to the thermal stress dimension".

[0083] According to the second fusion weight, if the user is a device operator who prefers structured data, the second fusion weight is set to "structured data: unstructured data = 0.75:0.25", and the answer is generated with the first target knowledge piece as the main part and the second target knowledge piece as the auxiliary part, and the reference answer is output: "The commonly used material of device M is chemical substance X, and the early production process may cause defects, and the manufacturing process needs to pay attention to the thermal stress dimension".

[0084] It should be understood that the preset retrieval model is used in the retrieval stage to search in multiple knowledge bases in parallel, which not only improves the retrieval efficiency, but also optimizes the personalization and timeliness of the answer through the fusion weight mechanism: the first fusion weight adjusts the priority of the pieces within the same knowledge type based on the relevance and update time, ensuring that the latest and most relevant information is given priority; the second fusion weight dynamically adjusts the proportion of knowledge types according to the user group feature information to cater to the individual needs of users, enhancing the relevance and practicality of the answer.

[0085] In some embodiments, the method further comprises: obtaining user feedback information, the user feedback information comprising a question pair composed of an audited query question and a target reference answer; constructing three types of knowledge bases based on a plurality of question pairs; performing retrieval in the three types of knowledge bases based on a preset retrieval model to obtain a plurality of target question pairs related to the query question, and generating a reference answer for the current query question according to the target reference answer in the target question pair.

[0086] Among them, the user feedback information refers to the question and answer result evaluation data actively submitted by the user or collected by the system during the use of the question and answer system, which forms effective feedback after internal audit process confirmation, including the actual query question put forward by the user and the correct reference answer confirmed after auditing, which constitutes a question pair. Correspondingly, for example, Figure 3As shown, the semiconductor knowledge base also includes a third type of knowledge base, which is a database of verified question-answer pairs, used to support fast and accurate responses to similar questions.

[0087] In some embodiments, the user feedback information includes multiple sources. For example, it can come from the user's rating or correction suggestions for the system's answers during use, such as marking the answer as inaccurate and providing the correct content. For example, it can also come from the screening of high-quality and confirmed question-answer pairs by experts in the semiconductor field reviewing the system's output question-answer records.

[0088] It should be understood that when a new query question is input, the default retrieval model first performs matching in the third type of knowledge base, and finds a historical question pair similar to the current question by semantic similarity calculation. If a matching target question pair is found, the reference answer in it is directly output as the response. For example, a user once asked "how to reduce the generation of residual after photoresist development", and the confirmed answer was "match the developer according to the type of photoresist, and replace the developer regularly". This question-answer pair is stored in the third type of knowledge base. When a subsequent user asks "possible reasons for residual after photoresist development", the system recognizes the semantic similarity and outputs the reference answer based on the answer to this question pair, achieving efficient reuse and improving the response accuracy and efficiency of common questions.

[0089] In some embodiments, when a user input query question is received, path selection decisions are made based on the type characteristics of the knowledge base itself, which are specifically divided into two main processing branches: one branch is for the second type of knowledge base facing unstructured data, whose content forms mainly include but are not limited to text files, various unstructured documents, and other data lacking predefined models; the other branch is for the first type of knowledge base facing structured data, whose content forms mainly include knowledge graphs with explicit patterns or structured organization forms such as relational or non-relational databases. One or more knowledge bases can be selected by the user, or the knowledge base corresponding to the user's preferred knowledge type can be directly selected according to the user group, and the environment and parameters required for the initialization of the corresponding processing flow are completed after the selection.

[0090] For the selected second-type knowledge base, the system performs operations of embedding vectorization on the content data in Chinese text form. At the same time, based on the user input question, the system synchronously initiates two kinds of retrieval operations: one is vector similarity matching, which is applied to the dedicated vector database to obtain retrieval results based on semantic similarity; the other is keyword matching, which is applied to the full-text retrieval database to obtain retrieval results based on text literal matching. Next, the system integrates the results obtained by the above two independent retrieval operations through a preset weighted fusion algorithm to form a preliminary fusion retrieval result. In order to further improve the relevance and timeliness of the results, the system combines the time weighting algorithm considering the document time information and a special reordering algorithm to filter and reorder the priority of the preliminary fusion result. In order to optimize the quality of the context information input to the large language model, remove the noise and focus on the key information, the system finally applies a context compression algorithm to the reordered result, which integrates the ability of the large language model to extract and compress the original text.

[0091] For the selected first-type knowledge base, the system first performs operations of extracting key entities and their potential relationships from the user question. Subsequently, the system obtains information from the graph database using multiple query strategies: including performing similarity matching in the graph vector space after vectorizing the extracted entities (vector matching), attempting to automatically convert natural language questions into graph query languages (such as Cypher queries) and executing them (text to Cypher query), and applying preset graph template queries for specific query patterns to construct and execute queries (graph template query). The graph data results obtained from different query methods (usually including nodes and relationships), the system performs deduplication processing to avoid redundant information. Finally, the processed results are integrated into a unified retrieval result set through a weighted fusion algorithm.

[0092] It should be noted that the question and answer system provided by the embodiments of the present application integrates multiple knowledge sources, covering the first-type knowledge base of structured data and the second-type knowledge base of unstructured data, and has strong multi-modal retrieval capability. Therefore, in the retrieval stage, the same query question may retrieve multiple knowledge fragments that meet the relevance threshold from different knowledge bases, and these knowledge fragments may differ in technical details, applicable scope, or expression angle, such as providing multiple different process parameter values, or some for explaining the cause of the problem. To avoid information confusion, logical conflicts, or key details being ignored, before performing S107 to generate the reference answer, the retrieved knowledge fragments are classified according to their semantic topics, information types, or functional attributes, and knowledge fragments with similar content or belonging to the same logical category are classified into a category. Within each category, a first fusion weight is calculated, and the knowledge fragments are weighted and fused by combining the two fusion weights to generate at least one reference answer that is semantically complete and clearly sourced.

[0093] Especially for complex or open questions, the question and answer system does not force all information to be combined into a single answer, but outputs multiple reference answers that are semantically independent and complementary to each other, retaining the diversity and context integrity of the information. In the semiconductor field, different material or parameter selection often corresponds to a specific process node, equipment configuration or product demand, and separate presentation of answers can better meet the dual requirements of information integrity and answer accuracy in the semiconductor field. Users can make accurate judgments in combination with their own scenarios.

[0094] Please refer to Figure 4 , Figure 4 is a schematic flowchart of an answer output method based on a semiconductor knowledge base provided by the embodiment of the present application, as Figure 4 indicated, the embodiment of the present application provides an answer output method based on a semiconductor knowledge base.

[0095] S201, obtaining a semiconductor knowledge base, the semiconductor knowledge base including a first type of knowledge base corresponding to structured knowledge and a second type of knowledge base corresponding to unstructured knowledge. It should be understood that different knowledge base construction paths are adopted for structured and unstructured knowledge, which improves the retrievability and usability of knowledge, and improves the breadth of knowledge, thereby improving the coverage and accuracy of answers. For details, please refer to the foregoing related embodiments, which will not be described here.

[0096] S202, receiving a query question input by a user, performing retrieval in the first type of knowledge base and the second type of knowledge base based on a preset retrieval model, and obtaining a plurality of knowledge segments related to the query question. For details, please refer to the foregoing related embodiments, which will not be described here.

[0097] S203, generating at least one reference answer based on the plurality of knowledge segments, and evaluating the overall credibility of each reference answer according to the confidence of the reference knowledge segment used by each reference answer.

[0098] Specifically, the generation of the reference answer can refer to the foregoing embodiments or related prior art, and each reference answer is fused by one or more reference knowledge segments, and the overall credibility of each reference answer is obtained by weighted average based on the individual confidence of the referenced reference knowledge segment.

[0099] In some embodiments, the confidence of the reference knowledge segment can be evaluated from multiple dimensions, and the specific dimensions and weights can be dynamically adjusted according to the actual application scenario. For example, in a process parameter query scene that emphasizes data accuracy, the confidence can be determined according to the source type, the confidence of the knowledge segment from the first type of knowledge base can be set to be high, such as 0.85, and the confidence of the knowledge segment from the second type of knowledge base can be set to 0.5. Finally, the confidence of each reference knowledge segment is weighted and averaged to obtain the overall credibility of the reference answer.

[0100] In some embodiments, the method further comprises: determining a confidence degree of each reference knowledge piece based on a relevance of each knowledge piece to the query question, a matching degree of the knowledge piece to the group characteristic information, and an update time of each knowledge piece.

[0101] Specifically, the relevance of each knowledge piece to the query question is semantic relevance, which is output by a preset retrieval model when performing retrieval, and the closer the semantic association, the higher the relevance. The creation or last revision time (i.e., update time) of each knowledge piece is obtained, and the closer the update time to the current time, the higher the timeliness. The preferred knowledge type corresponding to the group characteristic information is obtained, and the matching degree of the knowledge piece of the preferred knowledge type is higher. Illustratively, the relevance of each knowledge piece to the query question, the matching degree of the knowledge piece to the group characteristic information, and the update time of each knowledge piece can be respectively mapped to multiple scores by a pre-set scoring rule, and the weighted average of the multiple scores can obtain the confidence degree of each reference knowledge piece. The specific weights of the weighted average can be flexibly set according to the requirements, for example, the relevance weight is 0.5, the matching degree is 0.3, and the timeliness is 0.2, and the final confidence degree is calculated.

[0102] S204, analyze the problem intention of the query question, and determine the risk level of the current question and answer according to the problem intention and the group characteristic information of the user.

[0103] Specifically, the problem intention refers to the purpose of the user to obtain information by raising the query question, which can be determined by matching the problem keywords or semantic analysis of the large language model according to the preset rule, and further, the risk level of the current question and answer is comprehensively judged in combination with the group characteristic information (such as the post permission of the user) of the user. For example, when the problem intention is "query the standard film thickness of a certain process layer" and the user is a research and development engineer, it belongs to regular information acquisition, and the risk level is determined as low risk; if the problem intention is "adjust the etching time to improve the production capacity" and the user is a process responsible person, the permission of which includes process change, the risk level is determined as high risk.

[0104] In some embodiments, the method comprises: performing semantic analysis on the query question to determine the problem intention category of each query question; determining the influence degree of the current question and answer on production operation according to the problem intention category; determining the implementable degree of the problem intention according to the group characteristic information; and mapping the current question and answer to the corresponding risk level according to the influence degree and the implementable degree, wherein the risk level is associated with a corresponding first preset threshold and / or a pre-step.

[0105] The influence degree refers to the negative influence degree that the technical content involved in the user query question may cause to production operation dimensions such as product quality, process stability, equipment safety or production continuity if the technical content is executed or used incorrectly. The implementable degree refers to whether the user group to which the user belongs has the actual permission to execute the operation of the question according to the identity characteristics (such as post, permission) of the user group, for example, a front-line operator usually has no right to modify process parameters, while a senior engineer or process responsible person has the corresponding permission.

[0106] Specifically, the query question is subjected to semantic analysis, the problem intention category to which the query question belongs is identified, the potential influence degree of the intention is evaluated, for example, “how to adjust the etching time” belongs to the production line control intention category and directly affects the product quality, and the influence degree is high; and “the basic principle of a process” is only for knowledge learning, and the influence degree is low. Meanwhile, the implementable degree is determined in combination with the group characteristic information, for example, the questioner is an intern operator, and the implementable degree of the adjustment of the key process parameters is low; if the questioner is a process responsible person, the implementable degree is high. Finally, the influence degree and the implementable degree are combined to determine and map to the corresponding risk level. For example, high influence and high implementable degree correspond to a high risk level; high influence and low implementable degree correspond to a medium risk; and low influence can be directly mapped to a low risk.

[0107] In some embodiments, each risk level is associated with a different first preset threshold value, which is a reference knowledge fragment overall credibility lower limit value bound with the risk level, and is used to determine whether an additional pre-step needs to be performed, for example, a higher first preset threshold value is set in a high-risk scenario.

[0108] In some embodiments, each risk level can be pre-associated with different pre-steps, which not only prevents serious consequences that may be caused by unreliable information being directly used, but also tries to salvage and utilize part of the doubtful answers through differentiated pre-processes, thereby avoiding the reduction of the practicality of the question and answer system due to excessive caution.

[0109] In some embodiments, the problem intention category includes at least one of a principle cognition intention, a research and development planning intention, a production line control intention, and a fault diagnosis intention.

[0110] Among them, the principle cognition type of intent refers to a query for understanding the basic working principle or theoretical basis of a certain technology, device or process, which usually does not involve specific operation or parameter adjustment, and is mainly used for knowledge learning, training or technical understanding. The R&D planning type of intent refers to a query in the technical pre-research, process development or product design stage to obtain information on different technical routes and material selection. The production line control type of intent refers to a query related to process parameter adjustment, equipment setting or process monitoring, which is directly related to actual production operation. The fault diagnosis type of intent refers to a query for positioning the root cause, finding historical cases or obtaining suggestions when encountering abnormal phenomena or equipment alarms.

[0111] In some embodiments, different problem intent categories can correspond to different impact levels. For example, the principle cognition type generally does not affect actual operation and can be set to a low impact level; the R&D planning type can affect the technical route and can be set to a medium impact level; the production line control type and the fault diagnosis type are directly related to the production process and may cause batch abnormalities or equipment damage if the suggestion is incorrect, and can be set to a high impact level.

[0112] In S205, the presentation form and / or output authority of each reference answer are adjusted according to the risk level and the overall credibility. The presentation form refers to the display method of the reference answer in the user interface, such as whether to highlight, add a warning, provide the source of the basis, etc. The output authority refers to whether the answer is directly visible, whether it needs to be displayed after review, or whether it is only for specific users to view.

[0113] In some embodiments, S205 further includes: when the overall credibility is lower than or equal to a first preset threshold, at least one of the following pre-step is executed according to the risk level: S2051, based on the emphasis identifier, marking the reference knowledge fragment in the source semiconductor knowledge; S2052, submitting the suspicious reference answer to an automatic verification queue, and if the verification is passed, generating a reference answer with a traceability label; the traceability label includes at least one of the knowledge type, the update time and the confidence level identifier of the knowledge fragment; S2053, not outputting the suspicious reference answer, and submitting the query question and the corresponding suspicious reference answer to an expert review queue to obtain and output the reference answer that passes the review or the reference answer revised by the expert.

[0114] Please refer to Figure 5 , Figure 5 is a schematic flowchart of an answer output mechanism under risk control provided by an embodiment of the present application. As shown in Figure 5 , first, the overall credibility of each reference answer and the risk level of the current question and answer are obtained based on the steps of the foregoing embodiments, and it is judged whether the overall credibility is lower than or equal to a first preset threshold. If yes, at least one pre-step is triggered for execution.

[0115] In step S2051, the user is enabled to quickly locate the original content on which the answer is based and to check the context and accuracy thereof autonomously by marking the reference knowledge fragment in the original semiconductor knowledge with an emphasis identifier (e.g. highlighting). The preceding step does not interrupt the output of the answer, and the reference answer is still normally presented, so that the risk prompt is realized with low system overhead, the judgment right is given to the user through visual guidance, the misuse of false information in low-risk scenarios is prevented, the practicability and response efficiency of the question-answering system are maximally maintained, and the system is applicable to scenarios with low risk and the user has certain professional discrimination ability.

[0116] In step S2052, the doubtful reference answer is submitted to an automatic verification queue for automatic review by preset rules (e.g. parameter range verification, logic consistency check, etc.) or a verification model. After verification, an answer with a traceability label is generated, and the knowledge type, update time and confidence level are explicitly marked to enhance the transparency of the information source. The preceding step introduces a certain response delay, but the accuracy of the answer can be improved without human intervention, the cost of human review in long-term operation is reduced, the safety is ensured, and the practicability of the system is still maintained.

[0117] The traceability label is used to simplify the prompt core information, and the specific label style can be flexibly set. For example, structured knowledge is marked with a blue label, unstructured knowledge is marked with a yellow label, and hybrid knowledge is marked with a two-color splicing label to intuitively prompt the dominant knowledge source; the update time is colored according to the time efficiency, for example, green for the last 30 days, yellow for 30-90 days, and orange for more than 90 days to reflect the timeliness of the dominant knowledge; the confidence level is divided according to the confidence and the pre-set confidence level division threshold, and different confidence levels correspond to different labels. For example, the three are combined and displayed in the associated area of the reference answer or the source semiconductor knowledge to realize clear and readable traceability prompt.

[0118] In step S2053, the doubtful answer is not directly output, but is submitted to an expert review queue for manual confirmation or correction by a domain expert, and the reference answer that passes the review or is corrected by the expert is obtained and output. The preceding step actively sacrifices the response efficiency and system immediacy to ensure the absolute accuracy and safety of high-risk problems (e.g. key process adjustment), and although the practicability is temporarily inhibited, the serious consequences such as line abnormality and batch rejection caused by false information are effectively avoided.

[0119] It should be understood that the three types of preceding steps form a progressive selection structure, step S2051 prioritizes efficiency and provides a prompt, S2052 improves accuracy with controllable delay, and S2053 prioritizes safety at the expense of response efficiency.

[0120] The hierarchical response mechanism prevents serious consequences that may be caused by directly using unreliable information, and salvages and utilizes part of the suspicious answers as much as possible through differentiated pre-processes, avoids reducing the practicality of the Q&A system due to excessive caution, returns part of the judgment right of the effectiveness of the answers to the users through mechanisms such as emphasis identification and traceability labels, controls the operation cost, and realizes efficient and flexible risk prompting and control. In this way, the balance between risk control and system practicality is realized, avoiding the reduction of system usability caused by "one-size-fits-all" blocking, and preventing production hazards caused by output efficiency.

[0121] In some embodiments, S2051 can be selected for a first risk level (such as a lower risk scenario); S2052 can be selected for a second risk level (such as a medium risk scenario), and for a verified suspicious reference answer, S2051 can be further executed; and S2053 can be selected for a third risk level (such as a higher risk scenario). The risk level of the first risk level is lower than that of the second risk level, and the risk level of the second risk level is lower than that of the third risk level.

[0122] In some embodiments, S205 further includes: when the overall credibility is higher than a first preset threshold, or after S2051 is executed, or after S2052 is executed, performing a step of: S2054, associating the reference answer with a source semiconductor knowledge of the reference knowledge segment as a synchronous output of the retrieval result.

[0123] Specifically, as shown in Figure 5 S2054 is executed when the overall credibility of the reference answer is higher than the first preset threshold, or the emphasis identification processing of S2051 is completed, or the automatic verification and addition of the traceability label of S2052 are passed. While outputting the answer, a traceable link between the reference answer and the original knowledge source is established, so that the user can view the specific knowledge segment and its context on which the answer is generated, meet the user's demand for self-verification and in-depth review, and is especially suitable for cross-verification before the user makes a decision, realizing transparent and traceable Q&A interaction. For example, a clickable reference mark or "view details" entry is added to the answer output interface, and the user can jump to the knowledge graph node, process document paragraph or technical report where the original knowledge segment is located after clicking.

[0124] In some embodiments, when the reference answer that passes the audit or is revised by an expert is output, S2054 can also be executed synchronously, so that the user can conveniently obtain the related knowledge of the current Q&A, and guarantee the expandability and information integrity of the Q&A.

[0125] It should be understood that in the knowledge retrieval and question answering process, the user can intuitively locate the source of the knowledge through the visualization module (such as highlighting key data segments or displaying associated graph paths), verify whether there is logical consistency between the key facts, conclusions in the answer and the original data source in the knowledge base, and then verify the correctness of the retrieval source and the generated content.

[0126] In some embodiments, the S2052 further comprises: performing provenance verification on the questionable reference answer, including: identifying key facts or conclusion statements in the questionable reference answer, verifying the logical consistency of the key facts and conclusions with the reference knowledge segments in semantics; if the provenance verification passes, performing compliance verification on the questionable reference answer based on a preset expert review rule library.

[0127] Specifically, the automatic verification queue needs to undergo two verifications, namely provenance verification and compliance verification. The provenance verification refers to tracing the source of the key information in the questionable reference answer and confirming whether it can find a semantically consistent supporting basis in the original reference knowledge segment. The compliance verification refers to judging whether the reference answer conforms to the domain specification according to the preset expert review rule library.

[0128] The natural language processing technology is used to identify the key facts or conclusion statements in the questionable reference answer, and it is determined whether there is a semantically matched description in the corresponding reference knowledge segment. If the contents are consistent, the provenance verification passes and enters the compliance verification stage.

[0129] The expert review rule library is called to perform compliance verification. The expert review rule library includes at least one of professional knowledge logic checking rules, industry compliance checking rules, and knowledge conflict detection rules. The professional knowledge logic checking rules are used to detect whether the answer has contradictions in the physical or process level in the semiconductor field. The industry compliance checking rules are used to detect whether the answer conforms to the general standards or enterprise specifications in the semiconductor manufacturing field. The knowledge conflict detection rules are used to detect whether the answer contradicts the high-confidence knowledge in the existing knowledge base (such as the three types of knowledge base).

[0130] In some embodiments, the method further comprises: if the provenance verification or the compliance verification fails, submitting the query question and the corresponding questionable reference answer to an expert review queue; and outputting the target reference answer that passes the review or is corrected by an expert as the retrieval result. In this way, when the automated verification fails or potential errors are found, human professional judgment is introduced to ensure the accuracy and compliance of the output answer and prevent high-risk error information from being misused.

[0131] In some embodiments, the target reference answer corrected by an expert can also undergo the review process of the automatic verification queue to further improve the accuracy of the answer.

[0132] In some embodiments, the method further comprises: when the query question is submitted to the expert review queue, forming a question pair of the query question and the target reference answer, the question pair being used to update the semiconductor knowledge base or the preset retrieval model.

[0133] Specifically, when the query question is submitted to the expert review queue and the target reference answer after being reviewed and approved or corrected, the query question is paired with the finally confirmed answer to form a high-quality question pair. After being structured, the question pair is stored in the three-type knowledge base as a reference for subsequent similar questions. At the same time, the question pair confirmed by the expert can also be used to continuously optimize the preset retrieval model, for example, by fine-tuning the semantic matching model to improve the recognition accuracy of similar questions or enhancing the understanding ability of the model to professional terms and ambiguous expressions. It should be understood that the question and answer system forms a closed-loop feedback mechanism in use: the question submitted to the expert review queue and its finally confirmed correct answer will be used to update the semiconductor knowledge base or optimize the retrieval model. Thus, the question and answer system has the ability of continuous learning and can continuously correct the knowledge blind spot and judgment deviation from actual interaction.

[0134] Referring to Figure 6 , Figure 6 is a schematic flowchart of a semiconductor knowledge base building method provided by an embodiment of the present application. As shown in Figure 6 , the overall building process is divided into two parallel paths of structured knowledge and unstructured knowledge. The process starts with creating a knowledge base space, first selects a knowledge structure type, and then processes the two types of data respectively. For structured knowledge, entity and relationship information is extracted based on LLM and soft and hard coding rules, a knowledge graph is constructed and Embedding vectorization processing is performed, a graph database is obtained, and finally the construction is completed through visual inspection, manual feedback correction and expert review. For unstructured knowledge, visual detection combined with multi-modal double model fusion analysis technology is used to extract multi-modal information, including layout structure and text content; the text is adaptively blocked to generate Embedding vectors, and the construction is completed through visual inspection, manual feedback correction and expert review. Both paths emphasize the combination of automatic processing and manual intervention to improve knowledge quality and retrievability.

[0135] Referring to Figure 7 , Figure 7 is a schematic flowchart of a semiconductor knowledge base retrieval method provided by an embodiment of the present application. As shown in Figure 7As shown, the overall retrieval process is also divided into two parallel paths of structured knowledge base and unstructured knowledge base. The process starts from selecting the knowledge base, after the user inputs the question, retrieval is performed in the selected knowledge base. For unstructured knowledge base, retrieval results are obtained from the corresponding database through vector matching and keyword matching, then weighted fusion is performed, and dimensions such as time factor are reordered to realize filtering based on the first fusion weight, then the context is compressed, and finally the LLM generates an answer combining the prompt, retrieval information and original question. For structured knowledge base, the question is preprocessed, such as entity relationship extraction, vector matching, Cypher query statement conversion, graph query is performed in the graph database, the retrieval results are weighted and fused, filtering based on the first fusion weight is realized, after removing the node relationship, the answer is generated by the LLM combining the Schema information and the prompt, and the logical correctness is ensured by visualizing the original node relationship. The output answer can be visually inspected for the original text source, finally corrected by manual feedback and confirmed by expert review, and the problems in the expert review queue and the final confirmed correct answers are used to optimize the database.

[0136] In some embodiments, the fusion retrieval results (i.e. knowledge fragments) finally generated by the unstructured path and / or the structured path are obtained, the retrieval results are processed by information compression, focusing on the core relevant fragments. The compressed retrieval context information is integrated and spliced with the original user input question and the system preset prompt word template to form a structured input content sent to the LLM. The LLM performs understanding, reasoning and content generation based on the integrated input, and outputs an initial answer text for the user question.

[0137] The embodiment of the present application further provides a method for building and applying a semiconductor knowledge base, which comprises the following steps: S301, acquiring semiconductor knowledge, wherein the semiconductor knowledge is divided into structured knowledge and unstructured knowledge based on knowledge types; S302, performing data association on the structured knowledge according to a general or specific semiconductor field knowledge structure based on a first preset model or a preset template, so as to build a first type of knowledge base; S303, performing data reorganization on the unstructured knowledge according to a layout of the unstructured knowledge based on a second preset visual model group, so as to build a second type of knowledge base; S304, receiving a query question input by a user, performing retrieval in the first type of knowledge base and the second type of knowledge base based on a preset retrieval model, and obtaining a plurality of knowledge segments related to the query question; S305, receiving a query question input by a user, performing retrieval in the first type of knowledge base and the second type of knowledge base based on a preset retrieval model, and obtaining a plurality of knowledge segments related to the query question; S306, generating at least one reference answer based on the plurality of knowledge segments, and evaluating the overall credibility of each reference answer according to the confidence of the reference knowledge segment used by each reference answer; S307, analyzing the problem intention of the query question, and determining the risk level of the current question and answer according to the problem intention and the group characteristic information of the user; and S308, adjusting the presentation form and / or output authority of each reference answer according to the risk level and the overall credibility. For specific implementation, reference can be made to the related embodiments described above, which will not be described here again.

[0138] Please refer to Figure 8 , Figure 8 is a structural schematic block diagram of a computer device provided by the embodiment of the present application. The computer device can be a terminal device or a server.

[0139] Exemplarily, the method described above can be implemented in the form of a computer program, which can run on the computer device as Figure 8 indicated.

[0140] As Figure 8 indicated, the computer device comprises a processor, a memory and a network interface connected through a system bus, wherein the memory can comprise a non-volatile storage medium and an internal memory.

[0141] The non-volatile storage medium can store an operating system and a computer program. The computer program comprises program instructions, which, when executed, can make the processor execute any one of the answer generation method based on the semiconductor knowledge base, the answer generation method based on the semiconductor knowledge base, the method for building and applying the semiconductor knowledge base, the answer generation method based on the semiconductor knowledge base, and the retrieval method of the semiconductor knowledge base.

[0142] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0143] The internal memory provides an environment for running a computer program in a non-volatile storage medium, which, when executed by the processor, can cause the processor to perform any one of the semiconductor knowledge base-based answer generation method, the semiconductor knowledge base building and application method, the semiconductor knowledge base-based answer generation method, the semiconductor knowledge base-based answer generation method, and the semiconductor knowledge base retrieval method.

[0144] The network interface is used for network communication, such as sending assigned tasks.

[0145] It should be understood that the processor can be a central processing unit (CPU), and the processor can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0146] In one embodiment, the processor is configured to run a computer program stored in the memory to perform the following steps: S101, obtaining semiconductor knowledge, the semiconductor knowledge is divided into structured knowledge and unstructured knowledge based on knowledge type; S102, based on a first preset model or a preset template, data correlation is performed on the structured knowledge according to general or specific semiconductor field knowledge structure to construct a type of knowledge base; S103, based on a second preset visual model group, data reorganization is performed on the unstructured knowledge according to the layout of the unstructured knowledge to construct a type of knowledge base; S104, receiving a query question input by a user, performing retrieval in the type of knowledge base and the type of knowledge base based on a preset retrieval model, obtaining a plurality of knowledge fragments related to the query question; S105, determining a first fusion weight between each knowledge fragment of the same knowledge type based on the relevance of each knowledge fragment to the query question and the update time of each knowledge fragment; S106, obtaining group feature information of the user, determining a preferred knowledge type of the user according to the group feature information, and determining a second fusion weight between different knowledge types according to the preferred knowledge type; S107, obtaining a reference answer based on the first fusion weight and the second fusion weight according to the plurality of knowledge fragments.

[0147] In one embodiment, the processor is configured to execute a computer program stored in the memory to implement the following steps: S201, obtaining a semiconductor knowledge base, the semiconductor knowledge base including a first type of knowledge base corresponding to structured knowledge and a second type of knowledge base corresponding to unstructured knowledge; S202, receiving a query question input by a user, performing retrieval in the first type of knowledge base and the second type of knowledge base based on a preset retrieval model, and obtaining a plurality of knowledge segments related to the query question; S203, generating at least one reference answer based on the plurality of knowledge segments, and evaluating an overall credibility of each reference answer according to a confidence degree of a reference knowledge segment used by each reference answer; S204, analyzing a problem intention of the query question, and determining a risk level of the current question and answer according to the problem intention and group characteristic information of the user; and S205, adjusting a presentation form and / or output authority of each reference answer according to the risk level and the overall credibility.

[0148] For example, the processor is configured to execute a computer program stored in the memory, and is further configured to implement the steps of the answer generation method based on a semiconductor knowledge base, the answer generation method based on a semiconductor knowledge base, the answer generation method based on a semiconductor knowledge base, the method for building and applying a semiconductor knowledge base, and the retrieval method of a semiconductor knowledge base provided in any of the embodiments of the present application, which will not be described herein.

[0149] In an embodiment of the present application, a computer readable storage medium is also provided, which stores a computer program including program instructions. When the processor executes the program instructions, the steps of any of the answer generation method based on a semiconductor knowledge base, the answer generation method based on a semiconductor knowledge base, the method for building and applying a semiconductor knowledge base, the answer generation method based on a semiconductor knowledge base, and the retrieval method of a semiconductor knowledge base provided in the embodiments of the present application are implemented.

[0150] The computer readable storage medium can be an internal storage unit of the computer device, such as a hard disk or a memory of the computer device. The computer readable storage medium can also be an external storage device of the computer device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc.

[0151] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for generating answers based on a semiconductor knowledge base, characterized in that, The method includes: S101, Acquire semiconductor knowledge, wherein the semiconductor knowledge is categorized into structured knowledge and unstructured knowledge based on knowledge type; S102, based on the first preset model or preset template, the structured knowledge is associated with data according to the general or specific semiconductor field knowledge structure to construct a knowledge base; S103, based on the second preset visual model group, the unstructured knowledge is reorganized according to the layout of the unstructured knowledge to construct a second type of knowledge base; S104, Receive the query question input by the user, perform a search in the first type of knowledge base and the second type of knowledge base based on the preset search model, and obtain several knowledge fragments related to the query question; S105, Based on the relevance of each knowledge fragment to the query question and the update time of each knowledge fragment, determine the first fusion weight between knowledge fragments of the same knowledge type; S106, Obtain user group characteristic information, determine the user's preferred knowledge type based on the group characteristic information, and determine the second fusion weight between different knowledge types based on the preferred knowledge type; S107, Based on the first fusion weight and the second fusion weight, a reference answer is obtained according to several knowledge fragments.

2. The method according to claim 1, characterized in that, S102 further includes: If the structured knowledge is preset standard data, the preset template corresponding to the standard data is called to parse the content of the structured knowledge and generate a type of data corresponding to the structured knowledge. If the structured knowledge is not the preset standard data, the first preset model is called to parse the structured knowledge and generate a type of data corresponding to the structured knowledge.

3. The method according to claim 1, characterized in that, The second preset visual model group includes a layout detection model and a multimodal recognition model; S103 includes: Based on the layout detection model, the unstructured knowledge is parsed to obtain several blocks and block layout data. Based on the multimodal recognition model, the content of blocks of different modalities is parsed to obtain the block content data of each block; Based on the block layout data, several block content data are spliced ​​and recombined to obtain two types of data corresponding to unstructured knowledge.

4. The method according to claim 2 or 3, characterized in that, The aforementioned knowledge base includes a knowledge graph database, and the other type of data is knowledge graph data; The two types of knowledge bases include a vector database and / or a full-text database; the method includes: vectorizing the two types of data and storing them in the vector database, and / or storing the two types of data in the full-text database.

5. The method according to claim 3, characterized in that, The block content data includes text data, formula data, and table data in text format; the multimodal recognition model includes: A text recognition model is used to identify and parse the block knowledge existing in the text state to obtain the text data; A formula recognition model is used to identify and parse the block knowledge existing in the formula state to obtain the formula data; A table recognition model is used to identify and parse the block knowledge existing in the table state to obtain the table data.

6. The method according to claim 1, characterized in that, S106 further includes: Obtain initial fusion weights, wherein the weight of the structured knowledge fragments in the initial fusion weights is higher than the weight of the unstructured knowledge fragments; The weight of the preferred knowledge type is increased in the initial fusion weight to update the initial fusion weight and obtain the second fusion weight.

7. The method according to claim 1, characterized in that, S107 further includes: Based on the first fusion weight, knowledge fragments of the same knowledge type are integrated and spliced ​​into the target knowledge fragment; The target knowledge fragments are weighted and fused based on the second fusion weight to generate the reference answer.

8. The method according to claim 1, characterized in that, The method further includes: Obtain user feedback information, which includes question pairs consisting of reviewed query questions and target reference answers; Three types of knowledge bases are constructed based on several questions; Based on a preset retrieval model, a retrieval is performed in the three types of knowledge bases to obtain several target question pairs related to the query question; Generate a reference answer for the current query question based on the target reference answer in the target question pair.

9. A computer device, characterized in that, The device includes: Memory, used to store computer programs; A processor for executing the computer program and, in executing the computer program, implementing the answer generation method based on a semiconductor knowledge base as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to implement the answer generation method based on a semiconductor knowledge base as described in any one of claims 1 to 8.

Images