Industry chain knowledge question and answer interaction method and system based on large model

Abstract

The application discloses an industry chain knowledge question and answer interaction method and system based on a large model, groups each process node contained in a target industry chain based on attribute weight values and continuity between process nodes and upstream and downstream relationships, so that the obtained node set contains process nodes that not only reflect the relationship between the upstream and downstream, but also can realize the solidification of the relationship between the process nodes based on the process sequence through the arrangement order between the node sets, and then realize the solidification of the knowledge structure of the target industry chain, reduce the phenomenon that the knowledge is too loose and it is difficult to find the logic and rules contained therein due to the weak knowledge structure, and provide conditions for the model to generate targeted and clear answers. The application is applicable to the application of electronic digital data processing technology in the field of question and answer interaction related technology.

Description

Technical Field

[0001] This application relates to the field of electronic digital data processing technology, and in particular to a method for digital computing or data processing specifically applicable to a particular application, specifically a supply chain knowledge question-and-answer interaction method and system based on a large model. Background Technology

[0002] The industrial chain exhibits significant economies of scale. Through specialization and collaboration, the industrial chain can significantly improve production efficiency; for example, smartphone manufacturing requires the cooperation of thousands of companies worldwide. Products continuously add value at each stage of the industrial chain. This makes the industrial chain an important and indispensable target for industrial development.

[0003] However, different industries have their own unique characteristics, making it difficult to simply copy their experiences. For example, traditional industries like washing machines have supply chains characterized by low knowledge density and strong knowledge structure. Emerging industries like display panels (e.g., LCD, OLED) have supply chains characterized by high knowledge density and weak knowledge structure. Simply copying a supply chain knowledge Q&A interactive system designed for traditional industries may result in less than expected performance.

[0004] For example, patent publication number CN112099743A, titled "Interactive System, Interactive Device and Interactive Method" (main classification number: G06F16 / 958), not only realizes human-computer interaction functions but also improves the user's interactive experience to a certain extent. On the one hand, this demonstrates the great potential of electronic digital data processing technology in the field of question-and-answer interaction; on the other hand, it also shows that there is still a broad prospect for technological expansion in this field. Summary of the Invention

[0005] This application provides a method and system for interactive question-and-answer questions on supply chain knowledge based on a large model, in order to at least partially solve the above-mentioned technical problems.

[0006] The embodiments of this application adopt the following technical solutions:

[0007] Firstly, embodiments of this application provide a supply chain knowledge question-answering interaction method based on a large model, the method comprising:

[0008] According to the process, the process nodes included in the target industrial chain are grouped to obtain several node sets arranged in sequence; such that there is at least one pair of adjacent process nodes located upstream and downstream respectively, placed in two adjacent node sets; and each node set has at least two process nodes that need to be executed continuously in the process, and the sum of the attribute weight values ​​of each process node in each node set is not greater than a preset first threshold weight value; the attribute weight value is used to characterize the degree of negative impact on the target industrial chain when a problem occurs in its respective process node;

[0009] When a user's question regarding the target industry chain is detected, a preset semantic analysis model is used to process the question to obtain a first processing result; the first processing result represents the target process node targeted by the question.

[0010] If the central tendency measure index obtained based on each process node in the available node set is greater than the preset second threshold weight value, then the available node set and the adjacent node set are both determined as the target node set; the available node set is the node set to which the target process node belongs.

[0011] Based on the knowledge corresponding to each process node contained in the target node set, a first answer to the question is generated.

[0012] In an optional embodiment of this specification, the method further includes:

[0013] The semantic analysis model is the BERT model; and / or,

[0014] The first answer is generated using a preset T5 model based on the knowledge corresponding to each process node contained in the target node set.

[0015] In an optional embodiment of this specification, the method further includes:

[0016] If the central tendency measure index obtained based on each process node in the available node set is not greater than the second threshold of the weight value, then the available node set is determined as the target node set.

[0017] In an optional embodiment of this specification, the method further includes:

[0018] Collect data on historical events during the historical operation of the target industry chain;

[0019] Based on the data of the historical events, an attribute weight value is determined for each of the process nodes, such that the attribute weight value is positively correlated with the frequency of the historical events.

[0020] In an optional embodiment of this specification, the method further includes:

[0021] When a user is detected asking consecutive questions about the target process node, the weight value is reduced by a first threshold.

[0022] Based on the reduced weight value and first threshold, the grouping is re-executed to update the node set;

[0023] The first answer is generated based on the updated set of nodes.

[0024] In an optional embodiment of this specification, the method further includes:

[0025] When a user is detected asking consecutive questions about the target process node, the weight value of the second threshold is reduced.

[0026] The target node set is redetermined based on the reduced weight value and the second threshold.

[0027] The first answer is generated based on the redefined set of target nodes.

[0028] In an optional embodiment of this specification, the method further includes:

[0029] Tag the process nodes based on historical events;

[0030] When a user asks a question about the target industry chain, historical events with a matching degree greater than a preset matching degree threshold are identified as target events.

[0031] The knowledge corresponding to the tags matching the target event in each process node included in the target node set is fused with the target event to generate a second answer to the question.

[0032] In an optional embodiment of this specification, the method further includes:

[0033] When a user's question is detected regarding an event at at least one of the process nodes, the weight value is reduced by a first threshold.

[0034] Based on the reduced weight value and first threshold, the grouping is re-executed to update the node set;

[0035] The first answer is generated based on the updated set of nodes.

[0036] In an optional embodiment of this specification, the method further includes:

[0037] The target industry chain is the display panel industry chain.

[0038] Secondly, embodiments of this application also provide a supply chain knowledge question-and-answer interactive system based on a large model, the system comprising:

[0039] The grouping module is configured to: group the process nodes of the target industry chain according to the process steps, resulting in several node sets arranged in sequence; such that there is at least one pair of adjacent process nodes located upstream and downstream respectively, placed in two adjacent node sets; and each node set contains at least two process nodes that need to be executed continuously in the process step, and the sum of the attribute weight values ​​of each process node in each node set is not greater than a preset first threshold weight value; the attribute weight value is used to characterize the degree of negative impact on the target industry chain when a problem occurs in its respective process node;

[0040] The first processing result generation module is configured to: when a user asks a question about the target industry chain, process the question using a preset semantic analysis model to obtain a first processing result; the first processing result represents the target process node targeted by the question.

[0041] The target node set determination module is configured as follows: if the central tendency measure index obtained based on each process node in the available node set is greater than a preset weight value second threshold, then the available node set and the node set adjacent to it are both determined as the target node set; the available node set is the node set to which the target process node belongs.

[0042] The first answer generation module is configured to generate a first answer to the question based on the knowledge corresponding to each process node contained in the target node set.

[0043] Thirdly, embodiments of this application also provide an electronic device, including:

[0044] Processor; and

[0045] A memory configured to store computer-executable instructions, which, when executed, cause the processor to perform the steps of the method described in the first aspect.

[0046] Fourthly, embodiments of this application also provide a computer-readable storage medium storing one or more programs that, when executed by an electronic device including multiple applications, cause the electronic device to perform the steps of the method described in the first aspect.

[0047] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects:

[0048] This application provides a large-scale model-based knowledge-based question-and-answer interaction method for the industrial chain. Based on attribute weights, the continuity between process nodes, and upstream and downstream relationships, it groups the process nodes within the target industrial chain. This results in a node set that not only reflects upstream and downstream relationships but also solidifies these relationships through the arrangement of the nodes within the process set, thereby solidifying the knowledge structure of the target industrial chain. This reduces the problem of loosely structured knowledge that is difficult to discern as logically or systematically, thus facilitating the generation of targeted and clear answers. Furthermore, this division of process node sets also helps to differentiate the knowledge involved in each process node across different node sets, reducing the likelihood of unclear answers due to excessively high knowledge density. This method enables the application of electronic digital data processing technology in the field of question-and-answer interaction. Attached Figure Description

[0049] Figure 1 A schematic diagram illustrating the process of the supply chain knowledge question-and-answer interaction method based on a large model provided in the embodiments of this specification;

[0050] Figure 2 This is a schematic diagram of the structure of an electronic device in an embodiment of this specification. Detailed Implementation

[0051] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Similar elements in different embodiments are referred to by related similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of the present application. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to the present application are not shown or described in the specification. This is to avoid obscuring the core parts of the present application with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0052] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.

[0053] The serial numbers assigned to components in this document, such as "first" and "second," are used only to distinguish the described objects and have no sequential or technical meaning. The terms "connection" and "linkage" used in this application, unless otherwise specified, include both direct and indirect connections (linkages).

[0054] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0055] like Figure 1 As shown, the supply chain knowledge question-and-answer interaction method based on a large model in this specification includes the following steps:

[0056] S100: According to the process, the process nodes included in the target industrial chain are grouped to obtain a set of nodes arranged in order.

[0057] The methods described in this specification are executed by a large-scale industry chain knowledge question-and-answer interactive system.

[0058] An industry chain refers to the value creation and transmission system formed by enterprises or organizations at each stage of a process centered around a core product or service, from raw material supply and manufacturing to final consumption and recycling, through technological and economic connections. Essentially, it is a systematic expression of the division of labor and cooperation within an industry, emphasizing the synergy and interdependence between upstream and downstream sectors. For ease of explanation, this manual uses the display panel industry chain as an example.

[0059] For example, the display panel industry chain can be divided into three major stages: front-end array process, mid-end cell process, and back-end module process, each stage containing multiple sub-process nodes. Specifically:

[0060] 1. The front-end array process (array substrate preparation) may include: substrate cleaning, thin film deposition (which may include: gate layer deposition, gate insulating layer deposition, active layer deposition, source / drain (S / D) layer deposition, protective insulating layer deposition, transparent pixel electrode (ITO) deposition), and photolithography (which may include: coating, exposure, development, etching, inspection and repair).

[0061] 2. Mid-stage Cell process (liquid crystal cell assembly) may include: color filter substrate (CF) preparation (which may include: black matrix (BM) preparation, color filter layer preparation, protective layer preparation, transparent conductive layer (ITO) preparation, columnar spacer (PS) preparation), alignment layer processing, liquid crystal cell assembly (which may include drop-fill (ODF), vacuum bonding, sealant curing), cutting and inspection.

[0062] 3. Back-end module manufacturing process (module assembly) may include: polarizer bonding, drive circuit bonding (may include: COG (Chip on Glass) bonding, FPC (Flexible Printed Circuit) bonding), backlight system assembly (may include: assembling light guide plate, LED light source, diffuser, brightness enhancement film), and complete assembly and testing (may include: assembling LCD modules with backlight, housing, touch screen and other components into a complete display, and performing electrical performance tests (such as brightness, contrast, response time), aging tests (48 hours of continuous lighting) and appearance inspection).

[0063] As for which of the aforementioned sub-process nodes can be used as "process nodes" in this specification, they can be selected based on actual needs. For example, the sub-process node of "complete machine assembly" has never occurred in history; or due to the impact of end-of-line supply chain relocation policies, the sub-process node of "complete machine assembly" is not in the target supply chain (but has been allocated to other countries or regions); or the knowledge required for the sub-process node of "complete machine assembly" is too low. In this case, the sub-process node of "complete machine assembly" may not be used as a "process node" in this specification.

[0064] Furthermore, the granularity of the "process node" division can be determined according to actual needs. For example, in one embodiment, "color filter substrate (CF) fabrication" is a separate process node; in another embodiment, "black matrix (BM) fabrication, color filter layer fabrication, protective layer fabrication, transparent conductive layer (ITO) fabrication, and columnar spacer (PS) fabrication" are each a separate process node. The node set in this specification is not strictly required to be static. In optional embodiments, when continuous user inquiries about a particular process node are detected, the granularity of the process node should be refined, the node set should be re-divided, and then subsequent steps should be executed.

[0065] The node sets in this specification are arranged not only by process nodes within each set according to process steps, but also by process steps between different node sets. Different node sets may contain different numbers of process nodes. Grouping must adhere to the following principles: there must be at least one pair of adjacent process nodes, located upstream and downstream respectively, placed in two adjacent node sets (each node set must satisfy this condition, allowing a single process node to appear in two adjacent node sets simultaneously); and each node set must contain at least two process nodes that need to be executed consecutively in the process step ("consecutive execution" means the process cannot be interrupted, with strict time limits. For example, the black matrix (BM) fabrication and color filter layer fabrication process nodes need to be executed consecutively. If the interval between them is too long, resulting in inventory cycles, it will lead to interface contamination, film performance degradation, and affect yield. This enables the characterization of the continuity between process nodes). The sum of the attribute weight values ​​of each process node in the node set (each attribute weight value of each process node is a normalized value, which is comparable horizontally) does not exceed a preset first threshold value (which can be an empirical value). The attribute weight value is a quantitative result of the abstraction and synthesis of multiple attributes of a process node. The attribute weight value can at least be used to characterize the degree of negative impact on the target industrial chain when a problem occurs in its respective process node (the specific "problem" here can be determined according to the actual situation, which at least causes the interruption of the continuity of the target industrial chain, or incurs unexpected costs to maintain the continuity of the target industrial chain. "Negative impact" is not only economic loss, but also includes a comprehensive impact on production efficiency, corporate reputation, and other aspects). In related technologies, technical means that can achieve this quantification purpose are applicable to this specification when conditions permit.

[0066] Because the industrial chain is highly continuous, its operation is not only constrained by the situation of a single process node, but also by the connection between process nodes. The method in this specification represents this connection relationship by grouping process nodes.

[0067] S102: When a user asks a question about the target industry chain, a preset semantic analysis model is used to process the question to obtain a first processing result.

[0068] The purpose of semantic analysis is to support accurate and efficient information interaction and decision-making by parsing and understanding semantic information in natural or formal language. Any model capable of semantic analysis in related technologies is applicable to this specification, where conditions permit.

[0069] The first processing result obtained in this step indicates which process node the query targets. The target process node may or may not be unique.

[0070] In an optional embodiment of this specification, the semantic analysis model is the BERT (Bidirectional Encoder Representations from Transformers) model. The BERT model primarily follows the following principles of natural language: Bidirectional Context Dependency: The BERT model captures bidirectional contextual information in natural language through a bidirectional Transformer encoder. Traditional models (such as GPT) only process unidirectional language order from left to right or right to left, while the Attention function in BERT's encoder is bidirectional, capable of considering the context on both sides of the target word simultaneously. For example, in the sentence "The bank is on the river," BERT can understand that "bank" could refer to either "riverbank" or "bank" through bidirectional context, rather than relying solely on information from a single direction. Self-Attention Mechanism: BERT's self-attention mechanism, based on Transformers, dynamically assigns weights and aggregates information by calculating the similarity between each position in the input sequence and all other positions. This mechanism enables the model to capture long-distance dependencies, such as associating pronouns with antecedents in complex sentences. The core of the self-attention mechanism is "dynamic focusing," that is, automatically adjusting the degree of attention to information at different positions according to task requirements. Dynamic Word Embedding Patterns: Unlike static word embedding models (such as Word2Vec), BERT's word embeddings are dynamically generated. The same word is assigned different vector representations in different contexts. For example, "apple" will generate different embedding vectors in the sentences "bitten by a python" and "is my favorite programming language." This dynamism stems from the model's integration of contextual information, making word meaning representations closer to actual semantics. Pre-training-Fine-tuning Knowledge Transfer Patterns: BERT adopts a "pre-training + fine-tuning" model: First, it performs self-supervised learning on massive amounts of unlabeled text to master the general rules of language (such as grammar and word meaning); then, it performs supervised fine-tuning on a small amount of labeled data to adapt to specific tasks (such as sentiment analysis). This model simulates the human cognitive process of "broad learning followed by specialized deepening," for example, building a language foundation through reading a large number of books, and then mastering writing skills through targeted practice. Task Adaptation and Generalization Patterns: BERT achieves cross-task knowledge transfer through a fine-tuning mechanism. The general language representations learned during its pre-training phase can be quickly adapted to diverse tasks such as text classification, named entity recognition, and question answering systems by adjusting the output layer. For example, in the medical field, BERT can be fine-tuned into a disease diagnosis model; in the financial field, it can be transformed into a public opinion analysis tool. This generalization ability stems from the model's capture of the deep structure of language, rather than simply memorizing specific task patterns.

[0071] S104: If the central tendency measure index obtained based on each process node in the available node set is greater than the preset weight value second threshold, then the available node set and the node set adjacent to it are both determined as the target node set.

[0072] The target node set represents the scope of knowledge that needs to be referenced when generating a subsequent answer. This effectively narrows down the knowledge scope of the answer, making it more targeted and less general. The available node set in this specification refers to the node set to which the target process node belongs. Technical means in related fields that can achieve the desired search effect are applicable to this specification, provided conditions permit.

[0073] Trend measures are core tools in statistics used to describe the central location of data distribution. They can include at least one of the following: numerical means (e.g., arithmetic mean, geometric mean, harmonic mean, etc.) and positional means (e.g., median, mode, etc.).

[0074] S106: Based on the knowledge corresponding to each process node contained in the target node set, generate a first answer to the question.

[0075] In related technologies, any model capable of generating answers based on limited knowledge is applicable to this specification. For example, fine-tuning domain-specific models involve adapting a general-purpose model (such as BERT or GPT) using domain-specific data (such as enterprise documents or industry reports) to fit a specific knowledge base. Another example is multimodal knowledge base models, which have strong scalability and can handle multimodal data such as text, images, and audio, such as extracting text and image information from product manuals to generate answers. Representative models include the Qwen3embedding series, which supports multimodal long text understanding.

[0076] To achieve excellent response generation, in an optional embodiment of this specification, the model architecture is also designed, in which the large model adopts a two-layer serial structure.

[0077] The first layer: Semantic analysis models (core functions: parsing input intent and extracting key information). The core of semantic analysis is to transform text input into structured "semantic representations" (such as intent labels, entity information, semantic vectors, etc.). Commonly used models fall into two categories: The first is lightweight and efficient (suitable for real-time scenarios), such as BERT-base / BERT-Chinese: pre-trained models based on Transformer, which can achieve "intent classification + entity recognition" through fine-tuning (such as extracting "question type" and "core entities" from user questions). It is a basic choice for semantic analysis and has strong compatibility. Another example is ERNIE: which enhances "knowledge fusion" (such as entity-level and phrase-level masking) on ​​the basis of BERT, making it more suitable for semantic parsing that requires combining common sense or domain knowledge (such as question analysis in the medical and legal fields). The second is complex semantic understanding models (suitable for deep analysis), such as RoBERTa-Large: which optimizes BERT's pre-training strategy (longer training time, larger batch size), has stronger semantic representation capabilities, and is suitable for scenarios that require accurate understanding of ambiguity and multiple intents (such as semantic disambiguation of vague questions). For example, SpanBERT is optimized for "text span analysis" (such as extracting the core modifiers in a question), making it suitable for scenarios that require precise positioning of "semantic focus" (such as extracting the "core demands of a question" in question answering).

[0078] The second layer: Answer generation model (core function: generating natural answers based on semantic representation). Answer generation requires combining the "semantic information" (such as intent, entity, semantic vector) output from the first layer to generate logically consistent text. Commonly used models are divided into two categories: one is controllable generation (suitable for precise question answering / task-based dialogue), such as T5 (Text-to-Text Transfer Transformer): unifying "text understanding and generation" into a "text-to-text" task, it can directly receive the "semantic tags + key information" output from the first layer (such as "intent: weather query, entity: Beijing, date: today"), generating structured answers (such as "Beijing's temperature today is 25℃, sunny turning cloudy"), with extremely strong compatibility and support for multiple languages. Another example is BART-base / Large: based on an "encoder-decoder" architecture, it excels at "text reconstruction and generation," and can combine the semantic vectors from the first layer to generate logically coherent short sentence answers (such as "Based on your question (semantic analysis results), the solution is..." in customer service dialogue), and supports controllable optimization (such as limiting answer length and avoiding redundancy). Secondly, there are fluent generative models (suitable for open-domain dialogue / long text generation), such as GPT-3.5-turbo (API version) / GPT-2: autoregressive generative models based on the Transformer decoder, which can receive the "semantic summary" output of the first layer (such as parsing user questions into "semantic vectors + core intent") to generate natural and fluent open-domain answers (such as casual conversations and explanatory answers to complex questions), suitable for scenarios with high requirements for "naturalness of language". Another example is PEGASUS: optimized for "summary generation". If the scenario is "generating summary-style answers based on semantic analysis results" (such as extracting core conclusions from document semantic analysis), PEGASUS can more accurately retain key information and generate concise summary-style answers.

[0079] T5 is a pre-trained language model based on the Transformer architecture. It learns rich linguistic knowledge and patterns through pre-training on large-scale text datasets (such as the C4 dataset). Its pre-training goal is based on a "denoising autoencoder" method, which corrupts the input text and trains the model to reconstruct the original text, enabling the model to learn the semantic and syntactic information of the text. After pre-training, T5 can be fine-tuned on task-specific datasets to further improve its performance on specific tasks. This means that the knowledge reserve of the T5 model mainly comes from pre-training data and fine-tuning data; if the samples do not contain knowledge in a certain aspect, the model may indeed be unable to learn that part. To achieve its answer generation function based on knowledge within a specific range (knowledge within the target node set), in a further optional embodiment, it can be combined with an external search mechanism, such as Retrieval-Augmented Generation (RAG) framework.

[0080] RAG (Recursive Aggregate Query) technology combines traditional information retrieval systems with generative models. Before providing an answer, the model retrieves relevant content from a specific knowledge base. These retrieval results are then provided as context to the language model to generate a more accurate and verifiable answer. Models like BERT can serve as embedding models within RAG, mapping text to a low-dimensional vector space for knowledge retrieval. The T5 model can also be combined with the retrieval module to enhance generative capabilities using information retrieved from a specific knowledge base. For example, in some experiments, T5-large is used as a retrieval evaluator to fine-tune the model, determining the relevance of the document context to the input query. The method described in this specification also limits the specific scope of knowledge, which helps improve the relevance of the generated answer.

[0081] The supply chain knowledge-based question-and-answer interaction method provided in this application, based on large-scale models, groups the process nodes within the target supply chain according to attribute weight values, the continuity between process nodes, and upstream and downstream relationships. This ensures that the resulting node sets not only reflect upstream and downstream relationships but also solidify these relationships based on the process sequence through the arrangement of the node sets. This solidifies the knowledge structure of the target supply chain, reducing the problem of loosely structured knowledge that makes it difficult to find underlying logic and patterns. This facilitates the generation of targeted and clear answers. Furthermore, this division of process node sets can also, to some extent, divide the knowledge related to each process node among the node sets, reducing the likelihood of unclear answers due to excessively high knowledge density. On the one hand, this demonstrates the significant potential of electronic digital data processing technology in the field of question-and-answer interaction; on the other hand, it also shows that technological exploration in this field has broad prospects for expansion.

[0082] In an optional embodiment of this specification, the attribute weight value can not only characterize the loss caused by the anomaly of its respective process node, but also characterize the probability of the process node experiencing an anomaly, thereby improving the comprehensiveness of the characterization of negative effects. In this embodiment, historical event data is collected during the historical operation of the target industrial chain; based on the historical event data, an attribute weight value is determined for each process node, such that the attribute weight value is positively correlated with the frequency of the historical event. Historical events in this specification refer to events that have a substantial negative impact on the target industrial chain, or that pose a risk to the operation or output of the target industrial chain. For example, a delay in the maintenance of a process node; an adjustment of process parameters for a process node; or an excessively long storage period for an intermediate product. The selection of historical events can be determined based on actual needs and expert experience. Historical event data is used to characterize historical events. Historical events can be caused by human factors or by unexpected factors.

[0083] Considering that user interactions with the interactive system may not be instantaneous, and the system cannot effectively answer user questions with a single interaction for every task, the reasons for which may be multifaceted, in an optional embodiment of this specification to improve user experience, when continuous questions from the user regarding the target process node are detected (for a single task, multiple questions constitute continuous questions, indicating that the system's answers may not meet the user's needs, and semantic understanding ability is inherent to the model and cannot be improved in a short time, we can start by adjusting the knowledge scope, improving the ability to filter effective information, and improving the ability of the answer to represent information), the weight value first threshold is reduced; based on the reduced weight value first threshold, grouping is re-executed to update the node set. Reducing the weight value first threshold can reduce the number of process nodes in the node set, thereby reducing knowledge density. This results in a smaller knowledge scope for subsequent generated answers, and the matching degree between the answer and the question will increase accordingly, at least to some extent compensating for the "mismatch" phenomenon caused by semantic understanding ability. Appropriate answers can also be provided through subsequent interactions, provided the knowledge scope is accurate. After generating an answer based on the updated set of nodes, the task can continue to be executed by interacting with the user and adjusting the scope and depth of knowledge (adjusting the granularity of process nodes).

[0084] Furthermore, when a user is detected asking consecutive questions about the target process node (in an optional embodiment, consecutive questions are defined as the number of questions asked about a single task exceeding a preset threshold (an empirical value, greater than 2)), the weight value second threshold is reduced; based on the reduced weight value second threshold, the target node set is redefined. Considering that consecutive questions might also be due to a limited knowledge base, making it impossible to provide appropriate answers, in this embodiment, by reducing the weight value second threshold, the scope of knowledge is expanded without ignoring the correlation between knowledge points and the order of knowledge across upstream and downstream process nodes, thereby improving the effectiveness of providing answers to users.

[0085] In an optional embodiment of this specification, the accuracy and knowledge scope of the answer can be adjusted simultaneously. In this embodiment, not only a first answer but also a second answer is generated. First, tags are set for the process nodes based on historical events. When a user's question regarding the target industry chain is detected, a historical event with a matching degree greater than a preset matching degree threshold (an empirical value; in related technologies, any technical means used to determine the matching degree is applicable to this specification when conditions permit) is identified as the target event. The knowledge corresponding to the tags matching the target event in each process node included in the target node set is fused with the target event (fusion refers to analyzing the probability analysis and / or risk analysis of the occurrence of the historical event in the situation involved in the question. In related technologies, any technical means that can achieve this analysis is applicable to this specification when conditions permit), generating a second answer to the question. The second answer based on historical events naturally has higher accuracy, and unlike the first answer, it can also expand the knowledge scope.

[0086] Furthermore, when a user is detected asking a question about at least one of the process nodes (indicating that the question is highly targeted, the answer should also be highly targeted to meet the user's needs), the weight value first threshold is reduced; based on the reduced weight value first threshold, grouping is re-executed to update the node set.

[0087] In a further optional embodiment of this specification, the training process of the aforementioned two-layer cascaded large model is also designed. In this embodiment, the core of the training combining RAG, BERT, and T5 is modular training + end-to-end fine-tuning. First, the "retrieval" and "generation" capabilities are optimized separately, and then the two are made to work together through joint data. The following is a feasible training scheme.

[0088] The training objective is to have BERT handle "precise retrieval" (matching user questions with text from external knowledge bases), T5 handle "high-quality generation" (generating answers by combining retrieved knowledge), and ultimately, through collaborative training, enable "retrieval results" to efficiently support "generation needs," avoiding redundant retrieval or generating content detached from knowledge.

[0089] The core module and data preparation process includes: First, the core modules are divided into sections. The retrieval module's core is BERT. BERT (preferably BERT-Chinese or Sentence-BERT, the latter optimized for sentence vectors for faster retrieval) is used as a "text embedding model" to transform user queries and knowledge base text (documents, such as webpage fragments or document paragraphs) into low-dimensional vectors. The most relevant Top-K knowledge items are matched using vector similarity (e.g., cosine similarity). The generation module's core is T5. T5 (preferably T5-Chinese-base / large, adapted for Chinese scenarios) is used as a "knowledge augmentation generation model." The input is "user query + retrieved Top-K knowledge," and the output is a natural answer that matches the knowledge. The RAG framework connects retrieval and generation, providing "knowledge base management" (e.g., vector databases FAISS, Milvus) and "retrieval result filtering" logic to ensure that T5 only receives highly relevant knowledge.

[0090] Next, data preparation is required. Three types of data are needed, covering the entire process of "retrieval," "generation," and "collaboration." The first is retrieval training data: Query-Document paired data (e.g., "Question: Deposition time for color filter preparation" + "Document: Deposition time for color filter preparation under different process conditions is..."), with each data point labeled with a "relevance tag" (1 = relevant, 0 = irrelevant). The second is generation training data: Triple pairs of "Query + relevant Document, generate answer" (e.g., "Question + knowledge related to color filter preparation, generate deposition time for color filter preparation under different process conditions is..."), ensuring that the Document is the knowledge source for the answer (this can be manually labeled or using a public RAG dataset such as HotpotQA). The third is collaborative fine-tuning data: Real-world data of "Query, retrieved Top-K Documents, final answer" (including retrieval logs), containing negative examples of "low-relevance Documents causing generation errors," used to optimize the retrieval-generation collaboration.

[0091] The phased training scheme consists of three phases, ranging from independent to collaborative training. Phase 1 focuses on training the BERT retrieval module independently (goal: improving retrieval accuracy). This includes model initialization: using a pre-trained Sentence-BERT (or BERT-base) to avoid training from scratch and reduce computational costs. Next is task design: transforming retrieval into a "sentence-pair relevance classification task"—input is [CLS] Query [SEP] Document [SEP], output is a binary label of "relevant / irrelevant"; simultaneously, "vector similarity ranking" is trained (using Triplet Loss: ensuring the vector distance between the Query and relevant Documents is less than the distance between the Query and irrelevant Documents). Then comes the training parameters: the optimizer requires AdamW, with a learning rate of 2e-5 (a common learning rate for BERT-like models). The batch size is required to be 16-32 (adjusted according to GPU memory, e.g., 16 for 16GB of GPU memory). The training epochs are required to be 3-5, based on retrieval accuracy. Next is the use of data. Only "retrieve training data" is used. After training, the model is exported as an "embedded model" and connected to a vector database (such as FAISS) to test the Top-K retrieval effect for any query.

[0092] Phase 2 is used to train the T5 generation module separately (the training objective is to improve its knowledge integration ability). This includes model initialization: using a pre-trained T5-Chinese-base, whose pre-training objective is "text-to-text," naturally suited to tasks involving "input knowledge to output answer." Then comes task design: the input format is standardized to "Answer questions based on the following knowledge: {Document1} {Document2} ... {DocumentK} Question: {Query} Answer:", allowing T5 to learn to "extract information from knowledge and organize it into an answer" (essentially a "conditional generation task"). Next are the training parameters, including an optimizer requirement of AdamW, a learning rate of 1e-4 (the learning rate for T5 generation tasks is slightly higher than BERT). The batch size requirement is 8-16 (generation tasks consume more GPU memory; 16GB of GPU memory uses 8). The training cycle is 5-8 epochs, using BLEU (measuring the similarity between the answer and the reference text) and ROUGE-L (measuring logical coherence) as evaluation metrics. Training stops when BLEU ≥ 40. Next is data usage: only "generate training data" is used, and the number of input documents is fixed during training (e.g., K=3, that is, 3 relevant knowledge items are input each time), so that T5 can adapt to "multi-knowledge fragment fusion".

[0093] Phase 3 is used for end-to-end collaborative fine-tuning of RAG (the training goal is to adapt retrieval and generation). The models in the first two phases are "independently optimized," which may result in "retrieved knowledge not being needed for generation" (e.g., retrieving redundant information that T5 cannot efficiently utilize). Collaborative fine-tuning is needed to "align" the two. This includes framework construction: connecting the "BERT embedding model, vector database, and T5 generation model" into a complete RAG process (e.g., inputting a query, BERT generating query vectors, retrieving Top-K documents from the vector database, concatenating the query and Top-K documents into T5 input, and T5 generating the answer). Fine-tuning task design: without retraining all parameters of BERT and T5, only fine-tuning the "BERT output layer" (adjusting vector weights to make the retrieval results better fit the needs of T5) and the "T5 input layer" (optimizing the attention allocation to the retrieved knowledge). Introducing a "negative example penalty": If a retrieved document has low relevance to the query (labeled as 0), but T5 still uses it to generate an answer, a penalty term is added to T5's loss function (e.g., increasing the loss value by 10%), forcing T5 to "prioritize highly relevant knowledge." Training parameters: The optimizer requires AdamW, with a learning rate of 5e-5 (the learning rate is lower during fine-tuning to avoid disrupting existing capabilities). Batch size is required to be 8-12, and training epochs: 2-3 epochs (only a few epochs are needed for collaboration). Evaluation metrics: In addition to BLEU and ROUGE-L, a new "knowledge accuracy" metric is added (human or automatic judgment of whether the generated answer comes entirely from retrieved knowledge, without fabricated content), requiring a knowledge accuracy ≥ 90%.

[0094] Furthermore, this specification also provides a supply chain knowledge question-and-answer interactive system based on a large model, the system comprising:

[0095] The grouping module is configured to: group the process nodes of the target industry chain according to the process steps, resulting in several node sets arranged in sequence; such that there is at least one pair of adjacent process nodes located upstream and downstream respectively, placed in two adjacent node sets; and each node set contains at least two process nodes that need to be executed continuously in the process step, and the sum of the attribute weight values ​​of each process node in each node set is not greater than a preset first threshold weight value; the attribute weight value is used to characterize the degree of negative impact on the target industry chain when a problem occurs in its respective process node;

[0096] The first processing result generation module is configured to: when a user asks a question about the target industry chain, process the question using a preset semantic analysis model to obtain a first processing result; the first processing result represents the target process node targeted by the question.

[0097] The target node set determination module is configured as follows: if the central tendency measure index obtained based on each process node in the available node set is greater than a preset weight value second threshold, then the available node set and the node set adjacent to it are both determined as the target node set; the available node set is the node set to which the target process node belongs.

[0098] The first answer generation module is configured to generate a first answer to the question based on the knowledge corresponding to each process node contained in the target node set.

[0099] The system can execute the methods in any of the foregoing embodiments and achieve the same or similar technical effects, which will not be elaborated here.

[0100] Figure 2 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Please refer to it. Figure 2 At the hardware level, the electronic device includes a processor, and optionally also includes an internal bus, a network interface, and memory. The memory may include main memory, such as high-speed random-access memory (RAM), or non-volatile memory, such as at least one disk drive. Of course, the electronic device may also include other hardware required for other business operations.

[0101] The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. This bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 2 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0102] Memory is used to store programs. Specifically, programs may include program code, which includes computer operation instructions. Memory may include main memory and non-volatile memory, and provides instructions and data to the processor.

[0103] The processor reads the corresponding computer program from non-volatile memory into main memory and then runs it, forming a supply chain knowledge question-and-answer interaction system based on a large model at the logical level. The processor executes the program stored in memory and is specifically used to execute any of the aforementioned supply chain knowledge question-and-answer interaction methods based on the large model.

[0104] The above is as stated in this application. Figure 1 The supply chain knowledge question-and-answer interaction method based on a large model disclosed in the illustrated embodiments can be applied to a processor or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this application can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0105] The electronic device can also perform Figure 1 A knowledge-based question-and-answer interaction method for the industrial chain based on a large model was developed and implemented. Figure 1 The functions of the embodiments shown are not described in detail here.

[0106] This application also proposes a computer-readable storage medium that stores one or more programs, the programs including instructions that, when executed by an electronic device including multiple applications, perform any of the aforementioned large-model-based supply chain knowledge question-and-answer interaction methods.

[0107] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0108] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0109] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction set implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0110] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0111] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0112] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0113] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0114] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0115] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0116] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.

Claims

1. An industry chain knowledge question and answer interaction method based on a large model, characterized in that, The method comprises: According to the process, each process node contained in the target industrial chain is grouped to obtain a plurality of node sets arranged in order; so that at least one pair of adjacent process nodes located in the upstream and downstream are placed in two adjacent node sets respectively; and in each node set, there are at least two process nodes that need to be continuously executed on the process, and the sum of the attribute weight values of each process node in each node set is not greater than a preset weight value first threshold; the attribute weight value is used to represent the degree of negative impact on the target industrial chain in the case of problems in the process node to which it belongs; When detecting a user's question about the target industrial chain, a preset semantic analysis model is used to process the question to obtain a first processing result; the first processing result represents the target process node to which the question is directed; If the central tendency measure index obtained based on each process node in the available node set is greater than a preset weight value second threshold, the available node set and the node set adjacent thereto are both determined as a target node set; the available node set is the node set to which the target process node belongs; Based on the knowledge corresponding to each process node contained in the target node set, a first answer to the question is generated; The trend measure index is a core tool in statistics for describing the center position of data distribution, which can include at least one of the following: numerical average, position average; When detecting a user's continuous question about the target process node, the weight value first threshold is reduced; Based on the reduced weight value first threshold, the grouping is re-executed to update the node set; Based on the updated node set, the first answer is generated; When detecting a user's continuous question about the target process node, the weight value second threshold is reduced; Based on the reduced weight value second threshold, the target node set is re-determined; Based on the re-determined target node set, the first answer is generated.

2. The method of claim 1, wherein, The method further comprises: The semantic analysis model is a BERT model; and / or, The first answer is generated based on the knowledge corresponding to each process node contained in the target node set using a preset T5 model.

3. The method of claim 1, wherein, The method further comprises: If the central tendency measure index obtained based on each process node in the available node set is not greater than the weight value second threshold, the available node set is determined as the target node set.

4. The method of claim 1, wherein, The method further comprises: Collecting data of historical events in the historical operation process of the target industrial chain; Based on the data of the historical events, the attribute weight value of each process node is determined, so that the attribute weight value is positively correlated with the frequency of the historical event.

5. The method of claim 1, wherein, The method further comprises: Based on historical events, the process nodes are labeled; When detecting a user's question about the target industrial chain, a historical event with a matching degree greater than a preset matching degree threshold is determined as a target event. Fusing corresponding knowledge matching the target event in each process node included in the target node set with the target event, to generate a second answer to the question.

6. The method of claim 1, wherein, The method further comprises: When detecting a question of a user on an event of at least one process node, reducing the weight value first threshold; Based on the reduced weight value first threshold, re-executing grouping to update the node set; Based on the updated node set, generating the first answer.

7. The method of claim 1, wherein, The method further comprises: The target industrial chain is a display panel industrial chain.

8. An industry chain knowledge Q&A interaction system based on a large model, characterized in that, The system comprises: The grouping module is configured to: group each process node included in a target industrial chain according to a process, to obtain a plurality of node sets arranged in sequence; make at least one pair of adjacent process nodes located upstream and downstream respectively placed in adjacent two node sets; and in each node set, there are at least two process nodes that need to be continuously executed on the process, and the sum of attribute weight values of each process node in each node set is not greater than a preset weight value first threshold; the attribute weight value is used to represent the degree of negative impact on the target industrial chain in the case of problems in the process node to which it belongs; The first processing result generation module is configured to: when detecting a question of a user on the target industrial chain, use a preset semantic analysis model to process the question to obtain a first processing result; the first processing result represents a target process node to which the question is directed; The target node set determination module is configured to: if a concentration tendency measure index obtained based on each process node in an available node set is greater than a preset weight value second threshold, then determine the available node set and the node set adjacent thereto as target node sets; the available node set is a node set to which the target process node belongs; The first answer generation module is configured to: based on corresponding knowledge of each process node included in the target node set, generate a first answer to the question; The tendency measure index is a core tool in statistics for describing the position of the center of data distribution, and can include at least one of the following: numerical average, position average; When detecting a continuous question of a user on the target process node, reducing the weight value first threshold; Based on the reduced weight value first threshold, re-executing grouping to update the node set; Based on the updated node set, generating the first answer. When detecting a continuous question of a user on the target process node, reducing the weight value second threshold; Based on the reduced weight value second threshold, redetermining the target node set; Based on the redetermined target node set, generating the first answer.

Images