Enterprise knowledge learning method and device based on knowledge slices and intelligent dialogues

Abstract

The invention provides an enterprise knowledge learning method and device based on knowledge slicing and intelligent dialogue, and relates to the technical field of artificial intelligence, slicing is performed by fusing multi-dimensional features such as semantic similarity, content density, term concentration and structural integrity, and the knowledge slicing quality is improved; the fusion weight of BM25 reverse index and vector index retrieval is dynamically adjusted according to query features, so that self-adaption of fusion is realized, and the contradiction between recall and precision in different query scenes is solved; the reliability evaluation integrates the dimensionalities of coverage, evidence consistency, implication, evidence dispersion, generation uncertainty and the like, so that the answer reliability is improved; through knowledge point highlighting based on timestamps and AI accompanying learning of skipping, the learning efficiency of the user is improved; the personalized correlation score is dynamically updated based on the interactive behavior data and is used for a subsequent retrieval sorting process, so that accurate personalized recommendation is realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for enterprise knowledge learning based on knowledge slicing and intelligent dialogue. Background Technology

[0002] With the accelerated advancement of digital transformation, corporate training has become an important means to enhance organizational competitiveness. The current mainstream corporate training technology architecture mainly comprises three types of systems: online learning platforms that adopt a chapter-based content organization format and support standardized learning modes such as video-on-demand and live streaming; traditional knowledge base systems that store corporate knowledge assets based on a document tree structure and are typically equipped with keyword search functionality; and general-purpose intelligent question-and-answer robots that utilize natural language processing technology to achieve human-computer dialogue interaction.

[0003] The three technical solutions mentioned above all have significant limitations in practical applications: the smallest knowledge granularity of online learning platforms is usually a course unit or chapter module, making it difficult for employees to accurately locate the knowledge points corresponding to specific business scenarios; traditional FAQ (Frequently Asked Questions) systems only support rule-based keyword matching, making it difficult to understand semantically related questions such as "invoice approval process" and "expense reimbursement steps"; although third-party question-answering robots have conversational capabilities, they often make factual errors when answering questions in professional fields due to a lack of deep integration with the company's private data sources. These shortcomings result in low learning efficiency, insufficient knowledge accumulation, and difficulty in quantifying training effectiveness. Summary of the Invention

[0004] The purpose of this invention is to provide an enterprise knowledge learning method and apparatus based on knowledge slicing and intelligent dialogue, so as to at least solve one of the above-mentioned problems.

[0005] In a first aspect, the present invention provides an enterprise knowledge learning method based on knowledge slicing and intelligent dialogue, comprising: By combining semantic similarity, content density, terminology concentration, and structural integrity, the acquired enterprise knowledge resources are segmented to generate multiple knowledge slices, and a summary and question-answer pairs are generated for each knowledge slice. Each knowledge slice is stored in a structured format, and a keyword-based BM25 inverted index and a semantic-based vector index are constructed simultaneously. When a user's query request is received, the intent type corresponding to the query request is determined by intent recognition. When the intent type is a knowledge query intent, the fusion weight is dynamically calculated based on the query features in the query request. Based on the fusion weight, the merged candidate knowledge slices retrieved from the BM25 inverted index and vector index are fused and rearranged to obtain the target evidence fragment. The current answer is generated based on the target evidence fragment; the reliability of the current answer is assessed by combining coverage, evidence consistency, implication, evidence dispersion, and generation uncertainty; and the response data for the query request is returned based on the reliability assessment score. When users learn about audio and video resources related to response data through the player, a mapping relationship between the player's timestamp and knowledge slices is established to enable the highlighting and jumping of knowledge points based on timestamps; Collect user interaction data with knowledge slices, dynamically update the user's personalized relevance score for the corresponding knowledge slice based on the interaction data, and use it to influence subsequent search ranking results.

[0006] In an optional implementation, the acquired enterprise knowledge resources are segmented by combining semantic similarity, content density, terminology concentration, and structural integrity to generate multiple knowledge slices, including: Based on the preset window size, determine each candidate boundary in the enterprise knowledge resource and the text fragments on both sides thereon; Based on the text fragments on both sides of each candidate boundary, determine the semantic differences, content density changes, and term concentration changes on both sides of the candidate boundary; The multidimensional content boundary score of the candidate boundary is calculated by combining the semantic differences, content density changes, and term concentration changes on both sides of each candidate boundary with the structural integrity signal and fragment length corresponding to the candidate boundary. The target boundary is determined by a dynamic programming algorithm based on the multidimensional content boundary score and fragment length of each candidate boundary. The enterprise's knowledge resources are segmented according to the target boundaries to obtain multiple knowledge slices.

[0007] In an optional implementation, the fusion weight is dynamically calculated based on the query features in the query request, including: The fusion weight λ corresponding to the BM25 inverted index is calculated using the following formula: ; in, This represents the Sigmoid function. i 0、 i 1. i 2. i 3. i 4. i 5 indicates a learnable parameter or an offline calibration parameter. L q Indicates the query length. , Indicates the preset length scale, E q Represents the density of the substance. S spec Indicates rarity.A q Indicates ambiguity. Indicates a question.

[0008] In an optional implementation, the merged candidate knowledge slices retrieved from the BM25 inverted index and vector index are subjected to score fusion and rearrangement based on fusion weights to obtain the target evidence fragment, including: The recall depth of the BM25 inverted index and the recall depth of the vector index are determined based on the fusion weights. Based on the recall depth of the BM25 inverted index and the recall depth of the vector index, relevant knowledge slices of the query request are retrieved through the BM25 inverted index and the vector index respectively and then merged to obtain merged candidate knowledge slices. The fusion score of the merged candidate knowledge slices is calculated based on the fusion weight, and the target evidence fragment is determined from the merged candidate knowledge slices by cross-encoder rearrangement.

[0009] In an optional implementation, the reliability of the current answer is assessed by combining coverage, consistency of evidence, implication, dispersion of evidence, and generation uncertainty, including: The reliability assessment score is calculated using the following formula. CAR : ; in, COV This indicates the proportion of evidence covered by the answer. ENT This indicates the degree of logical implication of the evidence for the answer. CONS Indicating consistency of evidence, DISP Indicates the dispersion of evidence sources. CERT Indicates the generation confidence level. R len Indicates the length regularization term. len ans Indicates the actual length of the generated answer. L max This indicates the preset maximum length threshold. This represents the preset smoothing attenuation coefficient. α , β , c , d , e This indicates the preset weight.

[0010] In an optional implementation, response data for the query request is returned based on the reliability assessment score, including: When the reliability assessment score is higher than or equal to the preset assessment threshold, output the current answer and the reference information of the target evidence fragment; If the reliability assessment score is below the assessment threshold, only the citation information of the target evidence fragment is returned.

[0011] In an optional implementation, the user's personalized relevance score for the corresponding knowledge slice is dynamically updated based on interaction behavior data and used to influence subsequent search ranking results, including: Update users according to the following formula u With knowledge slices i Personalized relevance score: ; in, Indicates user u In time t Knowledge slices i Personalized relevance score, Indicates user u Prior to this update, knowledge slices i Historical relevance score, This represents the preset learning rate parameter. This represents the Sigmoid function. Indicates user u With knowledge slices i A collection of interactive events This indicates the preset behavior type weight. Represents the dwell function. d j Indicates interactive events j Duration of stay t j Indicates interactive events j Time, Indicates user u User-level time scale, Representing knowledge slices i global popularity r A coefficient representing global popularity. Indicates user u Knowledge slices i Context consistency score, Indicates user u In time t Previous and knowledge slices i Number of interactions , This indicates the preset adjustment parameters; Personalized relevance scores are used in subsequent retrieval ranking processes to be weighted and integrated with the baseline relevance scores obtained during the retrieval re-ranking stage, in order to update the ranking results.

[0012] Secondly, the present invention provides an enterprise knowledge learning device based on knowledge slicing and intelligent dialogue, comprising: The resource segmentation module is used to segment the acquired enterprise knowledge resources by combining semantic similarity, content density, terminology concentration and structural integrity, generate multiple knowledge slices, and generate a summary and question-answer pairs for each knowledge slice; The index management module is used to store each knowledge slice in a structured format and simultaneously build a keyword-based BM25 inverted index and a semantic-based vector index. The evidence retrieval module is used to determine the intent type of a query request when a user's query request is received by intent recognition; when the intent type is a knowledge query intent, it dynamically calculates the fusion weight based on the query features in the query request, and performs score fusion and rearrangement on the merged candidate knowledge slices retrieved from the BM25 inverted index and vector index based on the fusion weight to obtain the target evidence fragment; The answer feedback module is used to generate the current answer based on the target evidence fragment; to evaluate the reliability of the current answer by combining coverage, evidence consistency, implication, evidence dispersion and generation uncertainty; and to return response data for the query request based on the reliability evaluation score. The AI-assisted learning module is used to establish a mapping relationship between the player's timestamp and knowledge slices when users learn about audio and video resources related to the response data through the player, so as to realize the highlighting and jumping of knowledge points based on timestamps; The personalized recommendation module collects user interaction data with knowledge slices, dynamically updates the user's personalized relevance score to the corresponding knowledge slice based on the interaction data, and influences the subsequent search ranking results.

[0013] Thirdly, the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the enterprise knowledge learning method based on knowledge slicing and intelligent dialogue according to any of the foregoing embodiments.

[0014] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, performs the enterprise knowledge learning method based on knowledge slicing and intelligent dialogue as described in any of the foregoing embodiments.

[0015] The enterprise knowledge learning method and apparatus based on knowledge slicing and intelligent dialogue provided by this invention includes: segmenting acquired enterprise knowledge resources by combining semantic similarity, content density, terminology concentration, and structural integrity to generate multiple knowledge slices, and generating a summary and question-answer pairs for each knowledge slice; storing each knowledge slice in a structured format, and simultaneously constructing a keyword-based BM25 inverted index and a semantic-based vector index; when a user's query request is received, determining the intent type corresponding to the query request through intent recognition; when the intent type is a knowledge query intent, dynamically calculating the fusion weight based on the query features in the query request, and adjusting the fusion weight based on the BM25 inverted index and the vector index. The system retrieves candidate knowledge slices from the index, performs score fusion and rearrangement to obtain the target evidence slice; generates the current answer based on the target evidence slice; assesses the reliability of the current answer by combining coverage, evidence consistency, implication, evidence dispersion, and generation uncertainty; returns response data to the query request based on the reliability assessment score; establishes a mapping relationship between the player's timestamp and the knowledge slice when the user learns about the audio and video resources related to the response data through the player, enabling timestamp-based knowledge point highlighting and navigation; collects user interaction data with knowledge slices, dynamically updates the user's personalized relevance score to the corresponding knowledge slice based on the interaction data, and uses this data to influence subsequent search ranking results. This approach improves the quality of knowledge slicing by integrating multi-dimensional features such as semantic similarity, content density, terminology concentration, and structural integrity. It dynamically adjusts the fusion weights of the BM25 inverted index and vector index based on query features, achieving adaptive fusion and resolving the conflict between recall and precision in different query scenarios. Reliability assessment integrates dimensions such as coverage, evidence consistency, implication, evidence dispersion, and generation uncertainty, enhancing answer reliability. AI-assisted learning through timestamp-based knowledge point highlighting and navigation improves user learning efficiency. Finally, it achieves accurate personalized recommendations by dynamically updating personalized relevance scores based on interaction behavior data and using them in subsequent retrieval and ranking processes. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating an enterprise knowledge learning method based on knowledge slicing and intelligent dialogue, provided as an embodiment of the present invention; Figure 2This is a schematic diagram of a knowledge slice generation process provided in an embodiment of the present invention; Figure 3 A schematic diagram of the structure of an enterprise knowledge learning device based on knowledge slicing and intelligent dialogue provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Currently, corporate training mainly relies on online learning platforms, traditional knowledge bases, and general-purpose question-and-answer robots. Among these, online learning platforms lack fine-grained knowledge unit division, making it difficult for employees to quickly retrieve relevant content when encountering specific problems; traditional FAQs and keyword searches struggle to handle semantic relevance, resulting in low relevance of results; while general-purpose question-and-answer robots possess conversational capabilities, they lack integration with internal corporate knowledge and task processes, resulting in inaccurate and untraceable answers, failing to form a closed learning loop.

[0020] To address the shortcomings of existing enterprise learning and knowledge management scenarios, the embodiments of this invention mainly solve the following technical problems: 1. Fragmented and difficult-to-locate knowledge resources: Traditional systems cannot effectively identify the natural boundaries of knowledge units in documents and videos, resulting in coarse granularity of retrieval and citation, making it difficult for employees to quickly obtain the knowledge they need; 2. Rigid retrieval fusion method: The existing fixed weight fusion of BM25 and vector retrieval cannot adapt to different types of queries, resulting in low accuracy for short queries or entity-intensive queries, and insufficient recall for long queries or fuzzy queries; 3. Insufficient reliability assessment of answers: Common methods rely solely on coverage or simple similarity weighting, which cannot detect multi-source contradictions, insufficient logical implications, or uncertainties in generative models, and are prone to producing "illusionary" answers (i.e. answers that seem reasonable but are actually incorrect, fictitious, or inconsistent with known facts). 4. Personalized recommendation bias is serious: the existing click-based weighted method is prone to monopolizing popular items, ignoring the dynamic changes in user interests over time and behavioral differences, resulting in poor personalization effect; 5. Lack of a closed-loop learning process: The existing dialogue assistant is not linked with learning tasks, reminders, and companion learning functions, and cannot form a closed loop from "knowledge accumulation - intelligent retrieval - learning interaction - behavioral feedback".

[0021] To address the aforementioned issues, this invention provides an enterprise knowledge learning method and apparatus based on knowledge slicing and intelligent dialogue. Employing an end-to-end data flow and control flow design, it forms a complete intelligent learning closed loop encompassing resource parsing, knowledge slice generation, index construction, hybrid retrieval, answer generation and evaluation, interactive learning, personalized feedback, and secure billing. This loop aims to improve employees' knowledge acquisition efficiency, learning process management capabilities, and answer reliability within the learning platform.

[0022] To facilitate understanding of this embodiment, a detailed description of an enterprise knowledge learning method based on knowledge slicing and intelligent dialogue disclosed in this embodiment of the invention will be provided first.

[0023] This invention provides an enterprise knowledge learning method based on knowledge slicing and intelligent dialogue. This method can be executed by an electronic device with data processing capabilities and can be implemented by an enterprise learning assistant system. See also... Figure 1 The diagram shows a flowchart of an enterprise knowledge learning method based on knowledge slicing and intelligent dialogue. The method mainly includes the following steps S110 to S160: Step S110: Combining semantic similarity, content density, terminology concentration and structural integrity, the acquired enterprise knowledge resources are segmented to generate multiple knowledge slices, and a summary and question-answer pairs are generated for each knowledge slice.

[0024] In this embodiment of the invention, by analyzing enterprise documents or video content and comprehensively considering semantic changes, information density, distribution of professional terms and document structure, the content is intelligently segmented into knowledge fragments of appropriate size and clear themes. Then, a large language model is used to automatically generate content summaries and related question-and-answer pairs for each fragment, thereby transforming the complex knowledge resources into standardized knowledge slices that are easy to retrieve and learn.

[0025] Step S120: Store each knowledge slice in a structured format, and simultaneously construct a keyword-based BM25 inverted index and a semantic-based vector index.

[0026] In this embodiment, a standardized "digital archive" (structured storage) is created for each knowledge slice, and two different retrieval systems are established simultaneously: one is an "exact match" system (BM25 inverted index) for quickly finding slices containing specific keywords; the other is a "semantic understanding" system (vector index) for finding semantically similar slices based on the meaning of the question (rather than literal words). In this way, whether it is an exact query or a fuzzy question, the most relevant content can be retrieved quickly.

[0027] In practical implementation, knowledge slices can be stored in structured JSON format, including fields such as slice_id (slice ID, used to uniquely identify the knowledge slice), content (content, i.e., the main text of the slice), summary (summary, a brief overview automatically generated by the system or supplemented manually), qa_pairs (question-answer pairs, representing typical questions and corresponding answers associated with the slice), vector (vector, referring to the semantic representation of the slice after vectorization, used for similarity retrieval), labels (labels, representing the topic category or keyword annotation of the slice), and ts_range (timestamp range, mainly used for audio and video courseware, identifying the start and end times of the slice). Simultaneously, a BM25 inverted index and a vector index are established, and the permission information of tenants, departments, and users is written into the index layer. A hot / cold tiered storage approach can be adopted to improve query efficiency and resource utilization.

[0028] Step S130: When a user's query request is received, the intent type corresponding to the query request is determined by intent recognition; when the intent type is a knowledge query intent, the fusion weight is dynamically calculated based on the query features in the query request, and the merged candidate knowledge slices retrieved from the BM25 inverted index and vector index are fused and rearranged based on the fusion weight to obtain the target evidence fragment.

[0029] Once it is determined that the user's intent is to search for knowledge, the system will dynamically determine the weight of the two search methods based on the specific characteristics of the query (such as length, whether it contains technical terms, whether the question is ambiguous, etc.), and intelligently weight and merge the search results of the two search methods and reorder them, ultimately selecting the most relevant and reliable set of knowledge fragments as evidence to answer the question.

[0030] Step S140: Generate the current answer based on the target evidence fragment; assess the reliability of the current answer by combining coverage, evidence consistency, implication, evidence dispersion, and generation uncertainty; and return the response data for the query request based on the reliability assessment score.

[0031] This embodiment proposes a Comprehensive Answer Reliability assessment (CAR) model, which comprehensively considers coverage, evidence consistency, implication, evidence dispersion, and generation uncertainty. A preliminary answer can be generated using selected target evidence fragments. Then, a comprehensive assessment process is initiated, scoring the answer's reliability across five dimensions: whether the answer covers key evidence, whether the evidence is self-consistent, whether the evidence logically leads to the answer, whether the evidence sources are diverse, and whether the generation process is deterministic. Finally, the final response format is determined based on this score: highly reliable answers are returned directly, while less reliable answers are converted to providing only an evidence card or requiring user clarification, thus greatly avoiding providing incorrect or "illusory" information.

[0032] Step S150: When the user learns about the audio and video resources related to the response data through the player, a mapping relationship between the player's timestamp and the knowledge slice is established to realize the highlighting and jumping of knowledge points based on timestamps.

[0033] This embodiment provides an AI-assisted learning interaction function. The core data structure for establishing a timestamp mapping table, binding knowledge slices and video clips, and implementing the mapping relationship between the player's timestamps and knowledge slices can be: `chunk_id` refers to the knowledge slice ID, `ts_start` refers to the start timestamp, `ts_end` refers to the end timestamp, and `asset_id` refers to the resource ID. The player periodically reports the timestamps, and the system returns the corresponding knowledge point and highlights it. When a user clicks on a knowledge point, the player jumps to the corresponding location. Learning behavior data (dwell time, clicks, favorites) is fed back into the index for subsequent weighted optimization.

[0034] By establishing a "mapping table" between the timeline (timestamp) of audio and video and textual knowledge slices, when a user is learning in the player, the system can know in real time which knowledge point the currently playing content corresponds to (achieving highlighting). When the user clicks on a knowledge point, the player can automatically jump to the corresponding video segment for explanation, thus seamlessly connecting static knowledge cards with dynamic audiovisual content and achieving a precise learning experience of "what you click is what you learn, and what you learn is what you see".

[0035] Step S160: Collect user interaction behavior data with knowledge slices, dynamically update the user's personalized relevance score to the corresponding knowledge slice based on the interaction behavior data, and use it to influence the subsequent search ranking results.

[0036] In this embodiment, the user's learning behavior (such as clicking, saving, long dwell time, etc.) is continuously tracked, and these behaviors are regarded as "implicit feedback" on the user's interest in knowledge slices. Then, based on the type of behavior (such as saving has a greater weight than clicking) and the time of occurrence (recent behavior has a greater impact), the "personalized relevance score" between the user and these slices is dynamically adjusted. Finally, when the user searches again, this dynamically updated score will be merged with the basic relevance score, so that the content that the user is more interested in and more relevant will be prioritized and displayed, thereby realizing personalized knowledge recommendation.

[0037] The enterprise knowledge learning method based on knowledge slicing and intelligent dialogue provided in this invention improves the quality of knowledge slices by integrating multi-dimensional features such as semantic similarity, content density, terminology concentration, and structural integrity. It dynamically adjusts the fusion weights of the BM25 inverted index and vector index retrieval based on query features, achieving adaptive fusion and resolving the contradiction between recall and precision in different query scenarios. Reliability assessment integrates dimensions such as coverage, evidence consistency, implication, evidence dispersion, and generation uncertainty, improving answer reliability. AI-assisted learning through timestamp-based knowledge point highlighting and navigation enhances user learning efficiency. Finally, it achieves accurate personalized recommendations by dynamically updating personalized relevance scores based on interaction behavior data and using them in subsequent retrieval and ranking processes.

[0038] In some possible embodiments, see Figure 2 The diagram shown illustrates a knowledge slice generation process. Step S110 can be implemented by steps S210 to S250 as follows: Step S210: Based on the preset window size, determine each candidate boundary and the text fragments on both sides of it in the enterprise knowledge resource.

[0039] In this embodiment, a sliding window is used to scan the document: for each candidate segmentation point (i.e., candidate boundary) in the text, a segment of text to the left and right of the point is extracted according to the preset window size, forming two text segments to be compared, so that the difference between them can be calculated later to determine whether the point is suitable for segmentation.

[0040] Given candidate boundary positions i Define the left / right window: , (At the sentence / clause level) w It can be 3–5 sentences. For example, the window size can be set to 3 sentences, and the candidate boundary... i The text fragment on the left can be used to select candidate boundaries. i The previous three sentences, i.e., sentences [ i -2, i -1, iCandidate Boundary i The text fragment on the right can be used to select candidate boundaries. i The following three sentences, namely sentences [ i +1, i +2, i +3].

[0041] Step S220: Based on the text fragments on both sides of each candidate boundary, determine the semantic differences, content density changes, and term concentration changes on both sides of the candidate boundary.

[0042] Semantic differences can be obtained by calculating the cosine similarity of the overall semantic vectors of the two text segments; changes in content density can be obtained by comparing the difference in the average information entropy of the two text segments; and changes in terminology concentration can be obtained by statistically analyzing the differences in the concentration of domain-specific terms on both sides. Significant changes in these three dimensions collectively indicate that the location may be a natural topic or content boundary.

[0043] In one possible implementation, the aforementioned candidate boundaries i Semantic differences between the two sides It can be calculated using the following formula: , , ; in, The vector representation of the text segment on the left is the left window. L The average of all sentence vectors within the left window represents the overall semantics of the left window; The vector representation of the text segment on the right is the right-side window. R The average of all sentence vectors within the window represents the overall semantics of the right window; Represents the sentence vector.

[0044] Information density can be measured using the average word-by-word surprise / confusion level of a language model; based on this, candidate boundaries... i Content density changes on both sides It can be calculated using the following formula: ; ; in, S To represent a text segment, you can take... L , R ; D ( L The ) indicates the information density of the text segment on the left. D ( R The ) indicates the information density of the text segment on the right; Represents a text fragment SThe number of lexical units contained therein; t Words or subwords in a text segment are the smallest unit of computation in the text; history represents the historical context, referring to the context within the text segment. S In the middle, the current word element t All previous text content; Conditional probability refers to the probability that a language model predicts the next word based on its context history. t The probability of.

[0045] The above candidate boundaries i Terms concentration changes on both sides By analyzing the domain terminology set The IDF (Inverse Document Frequency) weighted density is obtained by calculating the density, and the calculation formula is as follows: , ; in, S To represent a text segment, you can take... L , R ; C ( L The ) indicates the terminology concentration of the text segment on the left. C ( R The ) indicates the terminology concentration of the text segment on the right; Represents a text fragment S The number of lexical units contained; IDF ( u ) indicates words u Inverse document frequency.

[0046] Step S230: Based on the semantic differences, content density changes, and term concentration changes on both sides of each candidate boundary, combined with the structural integrity signal and fragment length corresponding to the candidate boundary, the multidimensional content boundary score of the candidate boundary is calculated.

[0047] This invention proposes a Multi-Dimensional BoundaryScore (MBS) algorithm, which combines semantic similarity, content density, terminology concentration, and structural integrity to calculate the score of candidate boundaries. Subsequent steps involve segmentation based on high MBS scores, and the generation of summaries, knowledge points, and Q&As from a large model, which are then stored as JSON objects to form standardized knowledge slices. The segment length corresponding to the aforementioned candidate boundaries refers to the segment length from the previous boundary to the candidate boundary, such as the number of words or tokens.

[0048] The structural integrity signal corresponding to the candidate boundary can characterize respect for the layout / structure (headings, lists, tables, and chapter numbers, etc.). In other words, the structural integrity signal indicates that when determining whether to segment, it prioritizes respecting the inherent layout structure of the document itself. For example, it identifies natural dividing points such as before headings, after list items, table boundaries, or chapter numbers as more suitable segmentation boundaries, thereby avoiding forced cutting from unnatural places such as the middle of paragraphs and ensuring that the generated knowledge slices conform to human reading habits. Optionally, the candidate boundary... i Corresponding structural integrity signal score The calculation formula can be: ; in, Represents the Sigmoid function; Indicates the preset bias term; , , , These represent the pre-defined weight coefficients for different structural types, namely, heading weight, list weight, text block weight, and paragraph break weight. Indicates the title indicator, candidate boundary. i The value is 1 if a title exists at the specified location, otherwise it is 0. Indicates list indicator, candidate boundary i The value is 1 if there is a start or end position of a list item (such as a bullet point • or numbered list 1.), otherwise it is 0. Indicates text block indicator, candidate boundary i The value is 1 if there is a boundary of an independent text block (such as the beginning or end of a code block, chart, or citation); otherwise, it is 0. Indicates paragraph breakpoint indicator, candidate boundary i It takes 1 if it is a clear paragraph separator (such as a newline or paragraph indentation), otherwise it takes 0.

[0049] The aforementioned structural indicators (title indicators, list indicators, text block indicators, and paragraph break indicators, etc.) can be obtained through layout parsing / regular expressions (such as "Chapter N", "1)"). That is, in calculating structural integrity signals, in addition to semantic features, text layout formatting or explicit tags can be used as aids. Specifically, layout parsing techniques can be used to detect titles, lists, tables, and page breaks; for example, detecting bold fonts, font size variations, paragraph indentation, page breaks, and table / list structures to identify titles or entries; or regular expressions can be used to identify structural symbols such as "Chapter N", "1.", and "(I)". These explicit structural tags can significantly improve the system's accuracy in recognizing chapter and entry boundaries, thereby enhancing the rationality of segmentation points.

[0050] In one possible implementation, the calculation formula for the multidimensional content boundary scoring algorithm is as follows: ; in, Multidimensional content boundary score representing candidate boundaries; Represents the Sigmoid function; β 0 is the preset bias value. β 1. β 2. β 3. β 4. Weights set for learnable or experiential learning, for example, β 1. β 2. β 3. β The initial values ​​for 4 were 0.4, 0.2, 0.2, and 0.2, respectively. Subsequent A / B testing across different populations was used for continuous verification and optimization (A / B testing was conducted on different datasets to compare the differences in segmentation effects of different weight combinations, and adjustments were made based on experimental data feedback). β The value of Z(·) is used to obtain the weight configuration that is most effective in real-world applications; Z(·) represents robust standardization. This represents the length gate value, used to ensure that the segment length is appropriate.

[0051] Alternatively, the formula for robust standardization can be: ; Among them, median( x ) represents all candidate boundaries within the same document. x The median, MAD ( x ) represents all candidate boundaries within the same document. x The median absolute deviation. Robust MAD normalization of each dimension component of all candidate boundaries within the same document can suppress outliers.

[0052] Alternatively, the length gate function described above can be: ; in, This represents the segment length corresponding to the candidate boundary, i.e., from the previous boundary to... i The current segment length (number of words or number of tokens); This indicates a preset target segment length range, for example, L min =150 tokens L max =250 tokens.

[0053] In this embodiment, the boundary score is reduced when the segment length corresponding to the candidate boundary is insufficient. Specifically, at the candidate boundary... i At this point, if the length of the segment from the previous boundary to the current position is... Less than the preset minimum length threshold L min This indicates that the segmentation is too frequent or the fragments are too short. In this case, to avoid generating a large number of fragmented slices, a length gating function will be used. The score of this boundary is attenuated (weighted down). Even the candidate boundary i The semantic or information density has a trough. If the fragment length does not meet the minimum requirement, the probability of that position being selected as a segmentation point will also be reduced.

[0054] Step S240: Determine the target boundary using a dynamic programming algorithm based on the multidimensional content boundary score and fragment length of each candidate boundary.

[0055] This embodiment employs a dynamic programming algorithm to scan all candidate segmentation points (i.e., candidate boundaries) of the entire document, weighing the "boundary score" of each candidate segmentation point against the "fragment length" generated after segmentation, and finally selecting a globally optimal segmentation scheme. This ensures that the segmentation results can accurately capture content boundaries while guaranteeing that all generated knowledge slices are of uniform length and complete in content, avoiding overly fragmented or lengthy fragments.

[0056] The objective function of the above dynamic programming algorithm can be: ; in, B The target boundary set, i.e., the set of locations ultimately chosen as the splitting points; For the first k The length of each segment (in tokens or words); A preset target length (e.g., 150–250 tokens) is used to guide slice length balancing; The preset length penalty factor is used to control the trade-off between "maximizing boundary scores" and "fragment length balance".

[0057] The above dynamic programming algorithm can be implemented as follows: 1. Treat the document as a sequence and scan it from beginning to end; 2. At each candidate boundary location i Record "the previous segment's optimal score + the current boundary score - the length penalty" as the state transition equation; 3. Use DP (Dynamic Programming) to store the optimal partitioning result at the current position in a table; 4. Finally, the globally optimal boundary set B is found through backtracking.

[0058] This ensures that the total score is maximized while also achieving the desired result. Constraints are imposed to avoid generating unbalanced slices that are too long or too short.

[0059] Step S250: Divide the enterprise knowledge resources according to the target boundary to obtain multiple knowledge slices.

[0060] Based on the optimal splitting point (target boundary) determined by the dynamic programming algorithm, the original enterprise knowledge resources (such as long documents and videos) are cut into a series of independent and coherent knowledge fragments.

[0061] In this embodiment, considering the sensitivity of a single "semantic trough" to noise, content density (thematic shifts are often accompanied by sudden changes in information content) and terminology concentration (professional topic switching) are introduced to make boundary detection more robust. Structural integrity allows the algorithm to respect human document layout habits and reduce unnatural segmentation. MAD standardization and length gating avoid length imbalances and abnormal peak interference. Global DP is superior to point-by-point thresholding, ensuring overall segmentation quality. Thus, through multi-source evidence fusion, recall and accuracy are balanced; through robust standardization and gating, noise resistance is stronger and fragment granularity is more stable; and through global optimization, "jagged" segmentation caused by local optima can be avoided.

[0062] To facilitate understanding of step S130 above, the relevant intelligent question answering and skill arrangement will be introduced below.

[0063] This embodiment employs an intent recognition module for intent identification. Specifically, it can identify inputs as knowledge queries, data queries, plan recommendations, or learning enhancements based on a combination of classifiers and rules. This embodiment proposes a Query-Adaptive Dynamic Fusion (QADF) mechanism, which dynamically calculates the fusion weight λ for different queries. For short queries with dense entities, λ is increased, favoring the BM25 inverted index; for long queries or fuzzy questions, λ is decreased, favoring the semantic retrieval vector index. Cross-Encoder reordering refers to semantic matching and scoring of merged candidate knowledge slices, selecting, for example, Top-5 as target evidence fragments. Skill orchestration refers to calling corresponding tool modules through a decision tree to achieve data queries (points, learning time), plan recommendations (onboarding task push), and learning enhancement (task due reminders).

[0064] The following section provides a detailed introduction to the "Query Adaptive Dynamic Fusion (QADF)" mechanism of hybrid retrieval.

[0065] The fusion weight λ corresponding to the BM25 inverted index can be calculated using the following formula: ; in, This represents the Sigmoid function. i 0、 i 1. i 2. i 3. i 4. i 5 indicates a learnable parameter or an offline calibration parameter; L q Indicates the query length (number of words in the content). , This indicates a preset length scale (e.g., 6). The more short queries there are, the more the index will favor the BM25 inverted index. E q The entity density can be represented by the number of NER (Named-Entity Recognition) hits / number of words. The more proper nouns there are, the more it is biased towards the BM25 inverted index. S spec The rarity level can be represented by the average IDF or IDF variance. The rarer the index, the more it is biased towards the BM25 inverted index. A q This represents the ambiguity, with a value range of [0,1]. It can be estimated by a classifier / rule (for example, when there are interrogative sentences such as "how / why / please explain", the ambiguity is higher). This indicates an interrogative sentence; it takes the value 1 if the sentence is an interrogative sentence, and 0 otherwise.

[0066] The above formula indicates that short, entity-rich, and rare words result in reliable word-based matching (with a larger λ); while long, ambiguous, and highly interpretable words result in important semantic matching (with a smaller λ).

[0067] The aforementioned target evidence fragments can be determined as follows: The recall depth of the BM25 inverted index and the recall depth of the vector index are determined based on the fusion weights; based on the recall depths of the BM25 inverted index and the vector index, relevant knowledge fragments of the query request are retrieved through the BM25 inverted index and the vector index respectively, and then merged to obtain merged candidate knowledge fragments; the fusion score of the merged candidate knowledge fragments is calculated based on the fusion weights, and the target evidence fragments are determined from the merged candidate knowledge fragments through Cross-Encoder rearrangement.

[0068] This embodiment implements adaptive recall depth, and the recall size can be dynamically adjusted using the following formula: , ; in, This indicates the recall depth of the BM25 inverted index. Indicates the recall depth of the vector index. K0 represents the preset base recall depth, and λ represents the fusion weight. When there are many short entities, the recall depth of the BM25 inverted index is enhanced; conversely, the recall depth of the vector index is enhanced.

[0069] This dynamic adjustment of the fusion weight λ based on query characteristics avoids the bias of a fixed λ; it is linked to recall depth to improve the latency / effectiveness trade-off; and it can jointly learn the parameters in the fusion weight calculation formula with online feedback (clicks / satisfaction). i .

[0070] The above fusion score S It can be calculated using the following formula: ; Where λ represents the fusion weight of the BM25 inverted index, This represents the sparse matching score based on keywords. This represents a dense score based on vector semantic similarity.

[0071] Specifically, the retrieval results corresponding to the BM25 inverted index and the vector index are combined in the recall stage and then weighted and merged in the re-ranking stage. During the fusion, missing items are filled with the mean. This is the text matching score between the knowledge slice and the query under the BM25 inverted index, reflecting the degree of sparse matching between the query keywords and the slice documents in dimensions such as term frequency, inverse document frequency, and document length normalization; its calculation formula can be: ; in, Indicates query q With knowledge slices d Text matching score under the BM25 inverted index IDF ( t ) indicates a query q terms t Inverse document frequency, f ( t , d ) indicates the query terms t In knowledge slices d Frequency of occurrence in average d Represents the average document length of all knowledge slices, | d | Represents a knowledge slice d Length, k 1. b All of these are adjustable parameters.

[0072] It is the similarity score between knowledge slices and queries in the semantic vector space, and the calculation formula can be: ; in, Representing knowledge slices d With query q In semantic vector space, the similarity score is represented by cos(·), which denotes the cosine similarity function. E q and E d These are the vector representations of queries and knowledge slices after being processed by encoding models (such as SimCSE (SimpleContrastive Sentence Embedding), BERT (Bidirectional Encoder Representations from Transformers), E5 (Embeddings for Every Everything), etc.), where ||·|| represents the norm (modulus).

[0073] When a knowledge slice appears only in one retrieval channel, a missing completion mechanism is used to ensure the continuity of the fusion calculation. Specifically, for the first type of knowledge slice where only the BM25 inverted index exists for recall and there is no similarity score for the vector index, the completion strategy can be: take the average similarity score of the batch of retrieval results corresponding to the vector index as the similarity score corresponding to the first type of knowledge slice; or, take the minimum similarity score of 0.5 × Top-K, which is half of the lowest score among the batch of retrieval results corresponding to the vector index, to balance the semantic impact. For the second type of knowledge slice where only the vector index exists for recall and there is no text matching score under the BM25 inverted index, the completion strategy can be: take the average text matching score of the batch of retrieval results corresponding to the BM25 inverted index as the text matching score corresponding to the second type of knowledge slice. For the third type of knowledge slice where both the BM25 inverted index and the vector index are recalled, the weighted average is directly calculated according to the formula for calculating the fusion score.

[0074] To facilitate understanding of step S140 above, the reliability assessment of the answer will be described in detail below.

[0075] In some possible embodiments, the reliability assessment score can be calculated using the following formula. CAR : ; in, COV This indicates the proportion of evidence covered by the answer (i.e., coverage). ENT This indicates the degree of logical implication of the evidence for the answer. CONS Indicating consistency of evidence, DISP Indicates the dispersion of evidence sources. CERT Indicates the confidence level generated; R len This indicates a length regularization term used to apply a light penalty to excessively long answers (to prevent "word piling" from masking uncertainty). len ans This indicates the actual length of the generated answer (usually represented by the number of tokens). L max This indicates the preset maximum length threshold (which can be set based on experience with the task scenario, for example, 200~300 tokens). This represents the preset smoothing attenuation coefficient (used to control the intensity of the penalty; the larger the value, the milder the penalty; the smaller the value, the more severe the penalty). α , β , c , d , e This represents the preset weights, for example, α =0.25, β =0.25, c =0.2, d =0.15, e =0.15.

[0076] For coverage COV , This could be the percentage of quoted sentences / key statements in the answer; or the coverage percentage based on token alignment (based on cross-attention / retrieval hit alignment).

[0077] For consistency CONS , This is used to characterize the degree to which evidence does not contradict each other. Pairwise NLI (Natural Language Inference) can be performed on the evidence to determine the relationship between them, including entailment (i.e., the two pieces of evidence are mutually supportive), neutrality, and contradiction. Then, the contradiction rate (contradiction rate) is calculated. CONS =1-contradiction_rate, where the contradiction rate represents the proportion of all pairs of evidence that are judged to be contradictory.

[0078] For implied scores ENT , NLI can be used to assess the credibility of the evidence implied in the answer sentence. Since an answer usually consists of multiple sentences, the geometric mean or minimum value (a conservative approach) can be taken as the overall implied score of the entire answer.

[0079] For source tracing dispersion DISP , , can be the normalized entropy of the asset from which the evidence is derived, and its calculation formula can be as follows: ; Where p represents the proportion of each source. K Number of sources. H (p) represents the entropy value of the knowledge source distribution, used to measure the dispersion of knowledge sources (such as different knowledge slices, documents, or paragraphs) cited in the generated answer; it essentially reflects the diversity and balance of sources cited when generating the answer. If the generated answer relies on only a few sources (high concentration), then... H (p) is relatively small; if the answer cites multiple sources and is evenly distributed, then H ( The larger the value, the more dispersed and widespread the information sources are.

[0080] H (p) can be calculated as follows: Let there be a total of sources cited in the current answer. K Each source has several (e.g., different knowledge slices) and the citation ratio for each source is [missing information]. p i ,but: ; in, w i For the language model to generate answers, the first... i Attention weights or citation frequencies from each source are calculated using the following formula. H (p): ; This formula is used to measure the uniformity of the source distribution. Due to the maximum entropy... H max =ln K (When all sources are uniformly distributed), therefore DISP Using the above normalization form, and at the same time CONS Limit its effectiveness (multiple sources but contradictions do not add points).

[0081] for CERT , , is the confidence level during language model generation. It is expressed using normalized entropy. H norm To indicate uncertainty, then CERT =1- H norm . H normIt represents a measure of the uncertainty of the output distribution generated by a language model, calculated based on the entropy of the model's predicted probability distribution, and is used to measure the "confidence" or "certainty" of the model when generating the current answer.

[0082] H norm It can be calculated as follows: During the model generation process, the language model outputs a probability distribution for each token , y t represents the word to be predicted currently, which is at the t -th position in the sequence; y<t represents all the words that have been generated before the current position t ; x represents the input information of the model; represents that under the condition of the given input information x and the generated previous context y<t, the probability that the next word is y t . First, calculate the average token-level entropy through the following formula: ; where, T is the length of the answer. Then, according to the size V of the model's vocabulary, the maximum entropy is ln V , so the normalized entropy is: .

[0083] When the reliability evaluation score is higher than or equal to the preset evaluation threshold, output the citation information of the current answer and the target evidence segment; when the reliability evaluation score is lower than the evaluation threshold, only return the citation information of the target evidence segment. It should be noted that the evaluation threshold can be set according to actual needs and is not limited here.

[0084] In summary, using the geometric mean can sensitively penalize any extremely low value (that is, as long as there is a weak key dimension, the overall reliability decreases), which is better than the "mutual compensation" of linear weighting. Binding DISP and CONS to take the smaller value to avoid misjudging "multiple sources but mutually contradictory" as highly reliable. Introducing ENT and CERT can effectively suppress hallucinated answers. When CAR is lower than the preset evaluation threshold, only return the evidence card to avoid hallucinated answers.

[0085] To facilitate understanding of the above step S160, the personalized feedback weighting is introduced in detail below.

[0086] This embodiment proposes a Personalized Relevance Score (PRS), which can be used to update user information according to the following formula. u With knowledge slices i Personalized relevance score: ; in, Indicates user u In time t Knowledge slices i Personalized relevance score, Indicates user u Prior to this update, knowledge slices i Historical relevance score, This indicates the preset learning rate parameter (e.g., 0.05-0.2, used to control the fusion ratio of new and old information). This represents the Sigmoid function (used to normalize scores). Indicates user u With knowledge slices i A collection of interactive events This indicates the preset behavior type weight (used to assign different influences to different behaviors). This represents the dwell time function (used to weight the duration of user dwell time). d j Indicates interactive events j Duration of stay t j Indicates interactive events j Time, Indicates user u User-level timescales (e.g., 2–7 days, which can be adapted to user activity). Representing knowledge slices i global popularity r A coefficient representing global prevalence (i.e., debiasing strength). , Indicates user u Knowledge slices i The context consistency score (used to measure the similarity between the user's recently learned topics and the topics of this slice). Indicates user u In time t Previous and knowledge slices i The number of interactions (its reciprocal is used for exploratory weighting to avoid over-recommending learned slices). , This indicates the preset adjustment parameters (e.g., 0.05–0.1, used to control contextual consistency and the contribution of exploration items to the final score, respectively). The personalized relevance scores mentioned above are used to be weighted and integrated with the basic relevance scores obtained in the retrieval re-ranking stage during subsequent retrieval ranking, in order to update the ranking results.

[0087] The weighting for different behavior types can be set as follows: Favorite > Share > Click > Brief Dwell. A time decay function is introduced. It emphasizes recent behavior (which has a greater impact, while the impact gradually diminishes over time); it introduces global popularity to correct bias and avoid monopolization by popular trends; and it incorporates contextual consistency and exploratory items to ensure diverse recommendations. In search ranking, PRS The personalized relevance score is weighted and fused with the baseline score (i.e., the baseline relevance score), and the fusion strength is affected by... CAR (Reliability assessment score) adjustment ensures that personalization does not affect the reliability of the answer.

[0088] Specifically, the above events j Including behavioral types (Click, stay, save, share, complete), duration of stay d j and time t j Behavioral weights w a It can be set to: .

[0089] The dwell function can be (Insensitive to short stays, saturated with long stays), among which, d Indicates the duration of the stay. d Represents a preset characteristic time constant (e.g.) d =10s, dwell time at half saturation). Global popularity Used to de-bias popular items, for example, global_interactions i Representing knowledge slices i The total global interactions, where maxglobal represents the maximum global interactions. For example, =0.2, r =0.3, = =0.05.

[0090] During the retrieval re-ranking stage, the basic relevance scores of candidate knowledge slices and the user-personalized relevance scores are jointly weighted and fused. The final fused score is calculated using the following formula: ; min{1, CARi}; in, This represents the preset basic intervention weight, for example, =0.3. Base score S base,i (i.e., the above-mentioned fusion score) S (Reflecting knowledge slices) i Personalized scores are based on the semantic and textual relevance of the query. It reflects the degree of matching between a user's long-term learning preferences and interests within a specific time window. The system is based on reliability assessment factors. CAR i Dynamically adjust personalized intervention weights For low-reliability search results, reduce personalized intervention and avoid pushing users to "popular but inaccurate" content. This ensures search accuracy while introducing individualized ranking optimization, thereby improving the relevance of search results and user learning efficiency. (Based on final fusion score) S final,i Perform a descending sort and select the first few. N Each knowledge slice serves as the core source of evidence for generating answers.

[0091] In summary, this embodiment ensures that recent behaviors are more important through time decay, reflecting interest migration; it differentiates between "shallow interactions" and "strong intent" behaviors (collection / full viewing) through behavioral differentiation; it reduces the "Matthew effect" by popularization debiasing, allowing long-tail high-quality content to have a chance to be promoted; and it prevents overfitting to a single path by using contextual consistency and balancing the use / exploration of exploration items.

[0092] In practical applications, the following aspects should be noted: 1. Offline calibration + online self-calibration: MBS parameters β QADF parameters i You can first fit the data on manually labeled / replay data, and then use online feedback (satisfaction / reference clicks / completion rate) to make small-step updates.

[0093] 2. Explainable logs: Feature vectors and scores are written for each segmentation / retrieval / answer evaluation / personalization update, facilitating review and compliance auditing.

[0094] 3. Canary release: MBS → Segmentation, QADF → Fusion, CAR → Confidence Strategy, PRS → Sorting. It is recommended to advance the release in stages of 10% → 50% → 100% traffic, and monitor core KPIs (Key Performance Indicators), such as nDCG (normalized Discounted Cumulative Gain), coverage, follow-up inquiry rate, completion rate, and latency.

[0095] The four enhancement schemes above are decoupled from each other but can work together: MBS improves slice quality, QADF improves recall matching, CAR ensures answer reliability, and PRS personalizes the ranking but is constrained by CAR to prevent deviation.

[0096] In addition, the embodiments of the present invention also provide security and billing functions: AI point value billing can be used to finely calculate the resource consumption of ASR, slicing, vectorization, generation and other processes; knowledge security can be ensured through multi-tenant isolation and ABAC (Attribute-Based Access Control) permission control; and illegal content can be blocked by outputting sensitive words and compliance detection.

[0097] Compared with the prior art, the embodiments of the present invention have the following advantages: 1. Significantly improved knowledge slice quality: MBS integrates multi-dimensional features, making it more robust than traditional semantic valleys, and the granularity of the slice is reasonable.

[0098] 2. Retrieval Fusion Adaptive: QADF dynamically adjusts weights based on query features, resolving the conflict between recall and precision in different query scenarios.

[0099] 3. Higher reliability of answers: CAR comprehensively covers dimensions such as coverage, implication, and uncertainty, significantly reducing illusions and incorrect answers.

[0100] 4. More accurate personalized recommendations: PRS integrates time decay and behavioral differences, ensuring real-time performance while avoiding popular biases.

[0101] 5. Learning Interaction Loop: From knowledge accumulation → retrieval → dialogue → learning companionship → behavioral feedback, a closed loop is formed and continuously optimized.

[0102] 6. Scalable and controllable: Modular architecture, combined with security and billing mechanisms, supports large-scale enterprise deployment.

[0103] For ease of understanding, specific embodiments and their application scenarios are also provided in this invention.

[0104] Example 1: Knowledge Slice Generation and MBS.

[0105] The system performs sentence vector calculations on the uploaded documents; window size... w =3. Segmentation is triggered when the depth of the trough is greater than 0.3 and the segment length is between 200 and 500 characters. The process of determining the segmentation point uses a multidimensional content boundary scoring algorithm to calculate the score of the candidate boundary, and then uses a DP algorithm for global optimization to select the optimal segmentation point, generate knowledge slices, and call the large model to output a summary, knowledge points, and Q&A.

[0106] Application Scenario: When an enterprise administrator uploads a "Pricing Strategy Guide," the system automatically breaks it down into short segments of 200-300 words each and generates a summary and Q&A for each segment. Employee A can directly click on the "Pricing Formula" knowledge point to jump to the corresponding paragraph without having to read the entire 50-page document.

[0107] Example 2: Intelligent Question Answering and QADF.

[0108] The user queries "how to improve the sales process". The system identifies this as a knowledge search intent and calls both BM25 (Top-20) and vector index (Top-20). The fusion weight λ is dynamically calculated: longer queries with higher ambiguity result in smaller λ, favoring semantic retrieval. After merging the candidates, the data is input into a Cross-Encoder for rearrangement, ultimately selecting five evidence fragments to generate the answer.

[0109] Use Case: When salesperson B is preparing a client proposal, they input the question, "How to optimize the customer conversion process?". The system, relying more on semantic matching and using the QADF algorithm, returns five slices containing best practices, with each answer directly accompanied by a reference card. Salesperson B clicks on one of the references, and the system automatically jumps to the internal training video at 3 minutes and 25 seconds, providing a direct view of the explanation.

[0110] Example 3: CAR confidence assessment.

[0111] A given answer covers three pieces of evidence with the following parameters: COV=0.8, CONS=1.0, ENT=0.9, DISP=0.7, CERT=0.85. Substituting these parameters into the CAR calculation formula, we get CAR=0.87, which is higher than the threshold of 0.4. Therefore, the system directly outputs the answer.

[0112] Use Case: User C asks, "What is the calculation method for quarterly discounts?" The system finds two different document snippets as evidence, but CAR calculation confirms that the two are highly consistent and have sufficient coverage, so the answer is deemed reliable. User C receives a clear formula explanation and sees the source of the citation.

[0113] Example 4: AI-assisted learning and PRS.

[0114] The player's current time is t=120s, corresponding to chunk_id=c001, and the system highlights the relevant knowledge points. The user clicks "favorite," and this action is recorded. w save =1.0. Dwell duration function g_dwell=0.95, time decay factor. =3 days. The final PRS improved, and this slice ranked higher in subsequent searches.

[0115] Application Scenario: When User C watches a video and saves the knowledge point about "sales process optimization," the system automatically records this action and increases the ranking priority of related content in future searches. When he searches for "sales strategy" again a week later, the system will prioritize displaying the clips he previously saved.

[0116] Example 5: Security and Billing.

[0117] The system uses an AI-based point-based billing model: ASR is charged per minute, slices per page, and vectorization and generation are charged per token. Index write permission bits are used, and the retrieval stage is further filtered to ensure data isolation between different tenants. Output results are checked for sensitive words; inappropriate content is only displayed as a reference card.

[0118] Application Scenario: Training administrator D uploads a new policy document, and the system billing displays a consumption value of 300 points. Compliance checks detect sensitive data in a text file and automatically block its display to regular employees, making it visible only to administrators, thus ensuring compliance.

[0119] Corresponding to the above-described enterprise knowledge learning method based on knowledge slicing and intelligent dialogue, this invention also provides an enterprise knowledge learning device based on knowledge slicing and intelligent dialogue. See also... Figure 3 The diagram shows a structural schematic of an enterprise knowledge learning device based on knowledge slicing and intelligent dialogue. The device includes: The resource segmentation module 301 is used to segment the acquired enterprise knowledge resources by combining semantic similarity, content density, term concentration and structural integrity, generate multiple knowledge slices, and generate a summary and question-answer pairs for each knowledge slice; The index management module 302 is used to store each knowledge slice in a structured format and simultaneously build a keyword-based BM25 inverted index and a semantic-based vector index. The evidence retrieval module 303 is used to determine the intent type corresponding to the query request by intent recognition when a user's query request is received; when the intent type is a knowledge query intent, it dynamically calculates the fusion weight according to the query features in the query request, and performs score fusion and rearrangement on the merged candidate knowledge slices retrieved from the BM25 inverted index and vector index based on the fusion weight to obtain the target evidence fragment. The answer feedback module 304 is used to generate the current answer based on the target evidence fragment; to evaluate the reliability of the current answer by combining coverage, evidence consistency, implication, evidence dispersion and generation uncertainty; and to return response data for the query request based on the reliability evaluation score. The AI ​​learning companion module 305 is used to establish a mapping relationship between the player's timestamp and knowledge slices when the user learns about audio and video resources related to the response data through the player, so as to realize the highlighting and jumping of knowledge points based on timestamps; The personalized recommendation module 306 is used to collect user interaction behavior data with knowledge slices, dynamically update the user's personalized relevance score to the corresponding knowledge slice based on the interaction behavior data, and use it to influence the subsequent search ranking results.

[0120] The enterprise knowledge learning device based on knowledge slicing and intelligent dialogue provided in this invention improves the quality of knowledge slices by fusing multi-dimensional features such as semantic similarity, content density, terminology concentration, and structural integrity. It dynamically adjusts the fusion weights of the BM25 inverted index and vector index retrieval based on query features, achieving adaptive fusion and resolving the contradiction between recall and precision in different query scenarios. Reliability assessment integrates dimensions such as coverage, evidence consistency, implication, evidence dispersion, and generation uncertainty, improving answer reliability. AI-assisted learning through timestamp-based knowledge point highlighting and navigation enhances user learning efficiency. Finally, it achieves accurate personalized recommendations by dynamically updating personalized relevance scores based on interaction behavior data and using them in subsequent retrieval and ranking processes.

[0121] Furthermore, the resource segmentation module 301 is specifically used for: determining each candidate boundary and the text fragments on both sides of the enterprise knowledge resource according to a preset window size; determining the semantic differences, content density changes, and term concentration changes on both sides of the candidate boundary according to the text fragments on both sides of each candidate boundary; calculating the multidimensional content boundary score of the candidate boundary by combining the semantic differences, content density changes, and term concentration changes on both sides of each candidate boundary with the structural integrity signal and fragment length corresponding to the candidate boundary; determining the target boundary by using a dynamic programming algorithm based on the multidimensional content boundary score and fragment length of each candidate boundary; and segmenting the enterprise knowledge resource according to the target boundary to obtain multiple knowledge slices.

[0122] Furthermore, the aforementioned evidence retrieval module 303 is specifically used to: calculate the fusion weight λ corresponding to the BM25 inverted index using the following formula: ; in, This represents the Sigmoid function. i 0、 i 1. i 2. i 3. i 4. i 5 indicates a learnable parameter or an offline calibration parameter. L q Indicates the query length. , Indicates the preset length scale, E q Represents the density of the substance. S spec Indicates rarity. A q Indicates ambiguity. Indicates a question.

[0123] Furthermore, the aforementioned evidence retrieval module 303 is also used to: determine the recall depth of the BM25 inverted index and the recall depth of the vector index based on the fusion weight; retrieve relevant knowledge slices of the query request through the BM25 inverted index and the vector index respectively based on the recall depth of the BM25 inverted index and the recall depth of the vector index, and merge them to obtain merged candidate knowledge slices; calculate the fusion score of the merged candidate knowledge slices based on the fusion weight, and determine the target evidence fragment from the merged candidate knowledge slices through Cross-Encoder rearrangement.

[0124] Furthermore, the aforementioned answer feedback module 304 is specifically used to: calculate the reliability assessment score using the following formula. CAR : ; in, COV This indicates the proportion of evidence covered by the answer. ENT This indicates the degree of logical implication of the evidence for the answer. CONS Indicating consistency of evidence, DISP Indicates the dispersion of evidence sources. CERT Indicates the generation confidence level. R len Indicates the length regularization term. len ans Indicates the actual length of the generated answer. L max This indicates the preset maximum length threshold. This represents the preset smoothing attenuation coefficient. α , β , c , d , e This indicates the preset weight.

[0125] Furthermore, the above-mentioned answer feedback module 304 is also used to: output the current answer and the reference information of the target evidence fragment when the reliability assessment score is higher than or equal to the preset assessment threshold; and return only the reference information of the target evidence fragment when the reliability assessment score is lower than the assessment threshold.

[0126] Furthermore, the aforementioned personalized recommendation module 306 is specifically used to: update the user's profile according to the following formula. u With knowledge slices i Personalized relevance score: ; in, Indicates user u In time t Knowledge slices i Personalized relevance score, Indicates user u Prior to this update, knowledge slices i Historical relevance score, This represents the preset learning rate parameter. This represents the Sigmoid function. Indicates user u With knowledge slices i A collection of interactive events This indicates the preset behavior type weight. Represents the dwell function. d j Indicates interactive events j Duration of stay t j Indicates interactive events j Time, Indicates user u User-level time scale, Representing knowledge slices i global popularity r A coefficient representing global popularity. Indicates user u Knowledge slices i Context consistency score, Indicates user u In time t Previous and knowledge slices i Number of interactions , This indicates the preset adjustment parameters; Personalized relevance scores are used in subsequent retrieval ranking processes to be weighted and integrated with the baseline relevance scores obtained during the retrieval re-ranking stage, in order to update the ranking results.

[0127] The enterprise knowledge learning device based on knowledge slicing and intelligent dialogue provided in this embodiment has the same implementation principle and technical effects as the aforementioned enterprise knowledge learning method based on knowledge slicing and intelligent dialogue. For the sake of brevity, any parts not mentioned in the embodiment of the enterprise knowledge learning device based on knowledge slicing and intelligent dialogue can be referred to the corresponding content in the aforementioned enterprise knowledge learning method based on knowledge slicing and intelligent dialogue.

[0128] like Figure 4 As shown, an electronic device 400 provided in this embodiment of the invention includes: a processor 401, a memory 402, and a bus. The memory 402 stores a computer program that can run on the processor 401. When the electronic device 400 is running, the processor 401 and the memory 402 communicate through the bus, and the processor 401 executes the computer program to realize the above-mentioned enterprise knowledge learning method based on knowledge slicing and intelligent dialogue.

[0129] Specifically, the memory 402 and processor 401 mentioned above can be general-purpose memory and processor, without any specific limitations here.

[0130] This invention also provides a computer-readable storage medium storing a computer program. When a processor runs the computer program, it executes the enterprise knowledge learning method based on knowledge slicing and intelligent dialogue described in the preceding method embodiments. The computer-readable storage medium includes various media capable of storing program code, such as a USB flash drive, external hard drive, read-only memory (ROM), RAM, magnetic disk, or optical disk.

[0131] In this document, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A exists alone, A and B exist simultaneously, and B exists alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0132] In all examples shown and described herein, any specific values ​​should be interpreted as merely exemplary and not as limitations; therefore, other examples of exemplary embodiments may have different values.

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for enterprise knowledge learning based on knowledge slicing and intelligent dialogue, characterized in that, include: By combining semantic similarity, content density, terminology concentration, and structural integrity, the acquired enterprise knowledge resources are segmented to generate multiple knowledge slices, and a summary and question-answer pairs are generated for each knowledge slice. Each knowledge slice is stored in a structured format, and a keyword-based BM25 inverted index and a semantic-based vector index are constructed simultaneously. When a user's query request is received, the intent type corresponding to the query request is determined through intent recognition; When the intent type is a knowledge query intent, the fusion weight is dynamically calculated based on the query features in the query request. Based on the fusion weight, the merged candidate knowledge slices retrieved from the BM25 inverted index and the vector index are fused and rearranged to obtain the target evidence fragment. The current answer is generated based on the target evidence fragment; the reliability of the current answer is evaluated by combining coverage, evidence consistency, implication, evidence dispersion, and generation uncertainty; and response data for the query request is returned based on the reliability evaluation score. When the user learns about the audio and video resources related to the response data through the player, a mapping relationship between the player's timestamp and the knowledge slice is established to realize the highlighting and jumping of knowledge points based on the timestamp; Collect user interaction data with knowledge slices, dynamically update the user's personalized relevance score to the corresponding knowledge slice based on the interaction data, and use it to influence subsequent search ranking results.

2. The enterprise knowledge learning method based on knowledge slicing and intelligent dialogue according to claim 1, characterized in that, The method combines semantic similarity, content density, terminology concentration, and structural completeness to segment the acquired enterprise knowledge resources, generating multiple knowledge slices, including: Based on the preset window size, determine each candidate boundary and the text fragments on both sides of it in the enterprise knowledge resource; Based on the text fragments on both sides of each candidate boundary, determine the semantic differences, content density changes, and term concentration changes on both sides of the candidate boundary; The multidimensional content boundary score of each candidate boundary is calculated by combining the semantic differences, content density changes, and term concentration changes on both sides of each candidate boundary with the structural integrity signal and fragment length corresponding to the candidate boundary. Based on the multidimensional content boundary score and fragment length of each candidate boundary, the target boundary is determined by a dynamic programming algorithm. The enterprise knowledge resources are segmented according to the target boundary to obtain multiple knowledge slices.

3. The enterprise knowledge learning method based on knowledge slicing and intelligent dialogue according to claim 1, characterized in that, The step of dynamically calculating the fusion weight based on the query features in the query request includes: The fusion weight λ corresponding to the BM25 inverted index is calculated using the following formula: ； in, This represents the Sigmoid function. θ 0、 θ 1. θ 2. θ 3. θ 4. θ 5 indicates a learnable parameter or an offline calibration parameter. L q Indicates the query length. , Indicates the preset length scale, E q Represents the density of the substance. S spec Indicates rarity. A q Indicates ambiguity. Indicates a question.

4. The enterprise knowledge learning method based on knowledge slicing and intelligent dialogue according to claim 1, characterized in that, The process involves performing score fusion and rearrangement on the merged candidate knowledge slices retrieved from the BM25 inverted index and the vector index based on the fusion weights to obtain the target evidence fragment, including: The recall depth of the BM25 inverted index and the recall depth of the vector index are determined based on the fusion weights. Based on the recall depth of the BM25 inverted index and the recall depth of the vector index, relevant knowledge slices of the query request are retrieved through the BM25 inverted index and the vector index respectively and merged to obtain the merged candidate knowledge slices; The fusion score of the merged candidate knowledge slice is calculated based on the fusion weight, and the target evidence fragment is determined from the merged candidate knowledge slice by cross-encoder rearrangement.

5. The enterprise knowledge learning method based on knowledge slicing and intelligent dialogue according to claim 1, characterized in that, The reliability assessment of the current answer, which combines coverage, consistency of evidence, implication, dispersion of evidence, and generation uncertainty, includes: The reliability assessment score is calculated using the following formula. CAR : ； in, COV This indicates the proportion of evidence covered by the answer. ENT This indicates the degree of logical implication of the evidence for the answer. CONS Indicating consistency of evidence, DISP Indicates the dispersion of evidence sources. CERT This indicates the generation confidence level. R len Indicates the length regularization term. len ans Indicates the actual length of the generated answer. L max This indicates the preset maximum length threshold. This represents the preset smoothing attenuation coefficient. α , β , γ , δ , ε This indicates the preset weight.

6. The enterprise knowledge learning method based on knowledge slicing and intelligent dialogue according to claim 1, characterized in that, The step of returning response data for the query request based on the reliability assessment score includes: When the reliability assessment score is higher than or equal to a preset assessment threshold, the current answer and the reference information of the target evidence fragment are output. If the reliability assessment score is lower than the assessment threshold, only the reference information of the target evidence fragment is returned.

7. The enterprise knowledge learning method based on knowledge slicing and intelligent dialogue according to claim 1, characterized in that, The step of dynamically updating the user's personalized relevance score to the corresponding knowledge slice based on the interaction behavior data, and using this score to influence subsequent search ranking results, includes: Update users according to the following formula u With knowledge slices i Personalized relevance score: ； in, Indicates user u In time t Knowledge slices i Personalized relevance score, Indicates user u Prior to this update, knowledge slices i Historical relevance score, This represents the preset learning rate parameter. This represents the Sigmoid function. Indicates user u With knowledge slices i A collection of interactive events This indicates the preset behavior type weight. Represents the dwell function. d j Indicates interactive events j Duration of stay t j Indicates interactive events j Time, Indicates user u User-level time scale, Representing knowledge slices i global popularity ρ A coefficient representing global popularity. Indicates user u Knowledge slices i Context consistency score, Indicates user u In time t Previous and knowledge slices i Number of interactions , This indicates the preset adjustment parameters; The personalized relevance score is used to integrate with the basic relevance score obtained in the retrieval re-ranking stage according to confidence level during subsequent retrieval ranking, so as to update the ranking results.

8. An enterprise knowledge learning device based on knowledge slicing and intelligent dialogue, characterized in that, include: The resource segmentation module is used to segment the acquired enterprise knowledge resources by combining semantic similarity, content density, terminology concentration and structural integrity, generate multiple knowledge slices, and generate a summary and question-answer pairs for each knowledge slice; The index management module is used to store each knowledge slice in a structured format and simultaneously build a keyword-based BM25 inverted index and a semantic-based vector index. The evidence retrieval module is used to determine the intent type corresponding to the query request by intent recognition when a user's query request is received. When the intent type is a knowledge query intent, the fusion weight is dynamically calculated based on the query features in the query request. Based on the fusion weight, the merged candidate knowledge slices retrieved from the BM25 inverted index and the vector index are fused and rearranged to obtain the target evidence fragment. The answer feedback module is used to generate a current answer based on the target evidence fragment; to evaluate the reliability of the current answer by combining coverage, evidence consistency, implication, evidence dispersion, and generation uncertainty; and to return response data for the query request based on the reliability evaluation score. The AI-assisted learning module is used to establish a mapping relationship between the player's timestamp and knowledge slices when the user learns the audio and video resources related to the response data through the player, so as to realize the highlighting and jumping of knowledge points based on timestamps; The personalized recommendation module is used to collect the user's interaction behavior data with the knowledge slice, dynamically update the user's personalized relevance score to the corresponding knowledge slice based on the interaction behavior data, and use it to influence the subsequent search ranking results.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the enterprise knowledge learning method based on knowledge slicing and intelligent dialogue as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when executed by the processor, performs the enterprise knowledge learning method based on knowledge slicing and intelligent dialogue as described in any one of claims 1-7.

Images