Business English scene knowledge graph generation and intelligent retrieval method

Abstract

The invention discloses a business English scene knowledge graph generation and intelligent retrieval method, and relates to the technical field of natural language processing, and the method comprises the steps: constructing a document time sequence index and anaphora alignment mapping table, and registering an evidence anchor point index; extracting event units under the constraint and merging the event units into event clusters in a cross-document manner; arranging an event chain according to unified time and generating a process graph, and binding evidence anchor points to relation edges; analyzing the query into a query intention expression and a path constraint, and generating an evidence presentation sub-plan and an image-text combined execution plan; and positioning answers in the atlas, integrally returning the answers, paths and evidences, and meanwhile, forming replayable audit tracks. The method solves the problems of cross-channel time dislocation and confusion of references, and supports interpretable and auditable retrieval. The method is suitable for scenes such as mails, conference summary, instant communication and the like, guarantees locatable sources, path replay and evidence verification, and is convenient for internal control review.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a method for generating and intelligently retrieving knowledge graphs for business English scenarios. Background Technology

[0002] In the daily operations of large cross-border enterprises, activities such as order placement, contract negotiation, shipping notification, invoicing and reconciliation, and after-sales processing are scattered across three types of media: emails, meeting minutes, and instant messaging. The same facts often appear repeatedly by different people, in different languages, and with different expressions. Due to the mixing of server time, client time, and manual processing time, timestamps are inconsistent; due to the frequent use of aliases, abbreviations, and pronouns, the referents of participants and objects are diverse; and due to the fragmentation of channels, it is difficult to trace evidence back to the original text from the graph structure.

[0003] Existing technologies largely rely on keyword inversion, vector retrieval, and isolated entity recognition or event extraction, lacking cross-document common reference and a unified timeline. The generated knowledge graphs lack causal and state annotations, and the retrieval level tends to stack text fragments rather than providing path-level answers. Even when some systems construct process views, they fail to bind evidence locations to graph edges, resulting in unauditable answers and non-reproducible paths, making it difficult to meet compliance, review, and cross-team reuse requirements. Furthermore, parallel operations by teams in multiple time zones lead to intersecting timelines, with legal, sales, supply chain, and finance departments using their own tools, lacking a unified data framework. Existing solutions often rely on post-event manual review and offline spreadsheet splicing, which significantly increases the difficulty of misclassification, omissions, and accountability once the scale expands.

[0004] In a multi-source, asynchronous, and cross-domain business discourse environment, the lack of a unified time and stable reference coordinate system shared across the entire chain from discourse to retrieval makes it impossible to stably compile natural language queries into executable process graph paths and return auditable evidence. This problem typically occurs in procurement approval and delivery collaboration scenarios: pronouns are used to confirm approval in instant communication, aliases are used in emails to relay shipments, and project codes are used in meeting minutes to record completion dates for the following week. Due to the lack of document time sequence indexes and referential alignment mapping, event units are broken down, cross-document merging fails, and event chains are misordered, resulting in pseudo-concurrency or incorrect causality in the process graph. The retrieval end cannot use query intent expressions and path constraints to constrain searches, and can only return scattered fragments, which neither guarantees consistent answers nor aggregates evidence anchors. In scenarios with tight SLAs, cross-quarter settlements, or compliance spot checks, due to the lack of evidence presentation of sub-plans and combined text and graphics execution plans, the system cannot provide budget-compliant, highly diverse, and reproducible evidence packages and path-level explanations, leading to a significant decrease in the efficiency of internal control audits, customer inquiries, and dispute review. Misjudgments and delays further translate into delivery delays, penalty risks, and trust erosion.

[0005] At the same time, there is a lack of engineering compatibility between privacy and compliance: evidence cannot be shared externally without anonymization, and after anonymization, it loses its verifiability, breaking the audit chain. Summary of the Invention

[0006] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a method for generating and intelligently retrieving knowledge graphs for business English scenarios. This method involves constructing a document temporal index and a referential alignment mapping table, and registering evidence anchor point indexes. Under these constraints, event units are extracted and merged into event clusters across documents. Event chains are arranged according to a unified time frame to generate a process graph, with relationship edges bound to evidence anchor points. Queries are parsed into query intent expressions and path constraints, generating evidence presentation sub-plans and a combined text-image execution plan. The answer is located within the graph, and the answer, path, and evidence are returned in an integrated manner, simultaneously forming a replayable audit trail. This method resolves cross-channel time misalignment and referential confusion, supporting interpretable and auditable retrieval. It is applicable to scenarios such as emails, meeting minutes, and instant messaging, ensuring source locatability, path replay, and evidence verification, thus solving the technical problems described in the background section.

[0007] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: Methods for generating and intelligently retrieving knowledge graphs for business English scenarios, including: Perform turn-taking and time-sorting on emails, meeting minutes, and instant communications, generate document time-series indexes, simultaneously produce a referential alignment mapping table, unify pronouns, omissions, and aliases into stable references, and register evidence anchor point indexes; Under the constraints of the document temporal index and the reference alignment mapping table, the sentence segments are normalized into event units containing trigger words, participants, objects and time conditions, and event units pointing to the same change are merged across documents to form event clusters; The event clusters are sequentially arranged according to the document time sequence index and the causal relationships and state changes are marked. The sorting results are mapped to event nodes and relationship edges of the process graph, and the evidence anchor point corresponding to the original text position is recorded for each relationship edge. The user input is parsed into a query intent expression, and the roles, objects, and time requirements are translated into the process graph path constraints; the evidence presentation sub-plan is compiled according to the evidence requirements, and the evidence anchor index is used to generate a set for consumption in the execution phase; Using the query intent expression as the entry point, the answer that satisfies the path constraint is located in the process graph. The corresponding fragments of the evidence anchor point index are aggregated, and the answer, path and evidence anchor point are returned as a whole. The structured trajectory of the query and return is saved.

[0008] Furthermore, the talk round segmentation and time sorting include robust time alignment driven by cross-source anchors to estimate clock skew in each document and establish a unified timeline; Global optimization is performed using the boundary energy fused from semantic embedding changes, pause duration, and punctuation intensity. Dynamic programming and coordinate descent are used to solve the problem together, generating a set of turn boundaries and their unified timestamps, document identifiers, and paragraph offsets, which are then written into the document's temporal index.

[0009] Furthermore, the reference alignment mapping table is generated collaboratively by alias regularization and backreference parsing. Alias ​​regularization adopts optimal transmission with entropy regularization and introduces temporal and role priors. Backreference parsing combines semantic similarity, role compatibility, temporal proximity and process reachability. The priors derived from the process automaton and stable references are analyzed to search for candidate entities within a unified time window. Mentions that do not reach the threshold are marked for disambiguation and registered together with the original text character range and source document in the evidence anchor index.

[0010] Furthermore, the generation of the event unit adopts constraint structured decoding, including violations such as unique triggering of the same turn, participants and objects coming from the reference alignment mapping table, and time conditions being constrained by the document temporal index. The violations are quantized positively and participate in the unified solution. The time condition parsing maps absolute dates, relative periods, and corporate terminology to a unified time anchor, records the original character offset of the quadruple field and the source document, and associates the evidence anchor index to form an event unit object.

[0011] Furthermore, the cross-document coreference is softly assigned on the consistency graph, with priors consisting of stable referential consistency, triggering semantic similarity, and temporal proximity, and graph smoothing terms constraining the nearest neighbor event sharing assignment pattern; Event clusters are obtained through temperature annealing and threshold hardening. A cluster center vector and consistency score are calculated for each cluster. Isolated event units that fail hardening are returned to the event unit generation stage for re-evaluation and updating of the corresponding evidence anchor index.

[0012] Furthermore, the event chain construction introduces a global order variable and sorts the order by intervals. The set of hard constraints on the process limits the irreversible order and participates in the solution together with the weight matrix to obtain an acyclic order. Within the ordered space, directed pairs are selected based on causal posterior. The posterior features include rank difference, role leader matching, and evidence consistency. Nodes are mapped to discrete state labels using a state decoder. Adjacent high-confidence pairs are merged to form chain segments and summarized into the event chain set.

[0013] Furthermore, the process graph is constructed as a directed acyclic graph based on the candidate edge set, and each node is assigned a hierarchical variable and integerized during output. The inverse hierarchical loss is minimized by combining the prior hierarchy from the process stage and the margin is preserved. Edge-level evidence is aggregated by matching the edge semantic key vector with the anchor text vector according to the evidence anchor index. The edge attributes store the source document identifier, unified timestamp, paragraph offset, and the weight trajectory of the first few high-weight anchors.

[0014] Furthermore, the retrieval intent compilation generates a query intent expression by jointly decoding semantic parsing and graph priors. The parsing uses bundle search and includes violations such as time windows, role-object consistency, and reachability. The starting point, ending point, necessary and prohibited links are translated into path constraints. The path score integrates causal confidence, constraint satisfaction and time gap penalty, retains multiple candidates and outputs their respective constraint satisfaction and time smoothness indices.

[0015] Furthermore, the image-text joint execution plan selects evidence for candidate paths under the path constraints. The evidence presentation sub-plan characterizes diversity with a convex combination of semantic similarity kernel and temporal nearest neighbor kernel and adds importance weights, subject to evidence budget and terminal type limitations. Anchor points are extracted from the evidence anchor point index and the presentation order is determined to generate a plan object containing a path, a list of evidence, and presentation instructions.

[0016] Furthermore, the intelligent retrieval execution selects the optimal plan by jointly scoring the plan quality function and the entropy-temperature risk aggregation. The quality function includes three components: path coherence, time smoothing, and evidence credibility, and the risk temperature is configurable. The answer posterior is calculated for candidate endpoints in the neighborhood of the optimal path. The answer set for a single endpoint or multiple endpoints is selected by combining the threshold and the distinctness kernel control. The answer set is then bound to the path and evidence and written into storage.

[0017] Furthermore, the auditable return includes a chained audit trail and de-identified publication. The query intent expression, path constraints, optimal plan, answer set, document time sequence index, referential alignment mapping table and evidence anchor index are deterministically serialized in chronological order to generate a collision-resistant hash fingerprint. The serialization includes field sorting, time unit and numerical precision uniformity. The evidence vector is subject to sparse constraints on sensitive dimensions and non-sensitive metadata is retained, and a fixed format template is returned.

[0018] (III) Beneficial Effects This invention provides a method for generating and intelligently retrieving knowledge graphs for business English scenarios, which has the following beneficial effects: Based on document temporal indexing, referential alignment mapping tables, and evidence anchor indexes, cross-source time unification and referential normalization are achieved, fundamentally eliminating cross-timezone misalignment and name ambiguity. This ensures that subsequent extraction of the same object will not repeat or omit identification, and all locations can be replayed in the original text, significantly improving process stability. Constrained structured decoding is used to regularize sentence segments into event unit sets, which are then parsed using time conditions to fall onto a unified timeline. Entropy-temperature aggregation is then used to complete cross-document merging to form event clusters, cluster center vectors, and consistency scores. This compresses redundant expressions, solidifies consistent factual references, and maintains a two-way binding between facts and evidence anchors, enhancing cross-genre and cross-channel adaptability.

[0019] By introducing chain order rank and hard constraints of the process, a globally ordered set of event chains is obtained. Relationships and transitions are determined by causal posterior and state decoder. At the same time, process graphs are constructed by hierarchical assignment and acyclicity and evidence is organized according to edge evidence weight. Therefore, pseudo edges that are temporally adjacent but business-irrelevant are effectively suppressed, making causal paths clear, state semantics interpretable, and graph retrieval and playback consistent.

[0020] Semantic parsing and graph priors are jointly used to generate query intent expressions and path constraint sets. Path scoring integrates causal strength, temporal smoothness, and constraint satisfaction. Evidence presentation sub-plans are jointly selected based on diversity and importance and incorporated into the graph-text joint execution plan. Therefore, queries can directly fall into executable paths, the evidence package is compact and has high information density, and the returned results are both readable and easy to verify.

[0021] The system selects the optimal plan with risk aversion by using the plan quality function and path risk aggregation, and completes the single-endpoint or multi-endpoint decision by answer posterior. At the same time, it uses chained audit trails and hash fingerprints to solidify the entire process and cooperates with de-identified release. Therefore, it can still output robust answers and ensure traceability even in uncertain environments. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the process for generating and intelligently retrieving a knowledge graph for business English scenarios according to the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0024] Please see Figure 1 This invention provides a method for generating and intelligently retrieving knowledge graphs for business English scenarios, including: Addressing the multi-source, asynchronous, and cross-domain characteristics of business English communication (email, meeting minutes, and instant messaging), this project achieves discourse unification and referential alignment under a unified timeline constraint, thereby establishing an underlying structure that allows for chronological review, source identification, and reusable referentiality. This is achieved through the construction of a time-series index. Alignment Map with Reference This unifies the pronouns, omissions, and aliases of the same real object in different texts into a stable designation, thereby providing a coherent data framework for event extraction, event chain construction, and evidence anchor point backtracking.

[0025] Step 101: Under the conditions of device clock differences and language style differences in multi-source texts, construct a unified timeline and stable talk wheel boundaries to form a time-series index that can be consumed across documents. .

[0026] In business English contexts, emails often include server delivery timestamps, instant messaging records are typically based on the client's local time, and meeting minutes may be compiled by different participants in different locations, resulting in varying source timestamps. Inconsistency exists between them. Without prior unification, the segments of each source text will be misaligned due to clock drift, time zone crossings, and delays. Subsequent event unit extraction will break the same fact into different time slots, making it difficult to construct a reliable event chain.

[0027] Therefore, clock skew is estimated using cross-source time alignment. Then, the boundary set is determined by turn segmentation driven by both semantics and prosody. This generates a time-series index. This serves as the sole time reference for all subsequent extractions.

[0028] unreliable source timestamps Establishing a robust mapping between the data and sparse anchor facts is key to achieving a unified timeline. This is achieved by first extracting a unified timeline from cross-source high-confidence anchors (such as submitting a Purchase Order, signing an NDA, or receiving a shipping notification). Then check each document clock deviation Robust estimation is performed to minimize the deviation of the aligned sequence from the anchor point, while sparse regularization is introduced to suppress overfitting. This yields the document time offset estimate. : Where: anchor time : Originates from a unified time of a high-confidence event across documents; source timestamp :document Center and anchor point Corresponding original time; clock deviation :document The global time offset, Shift to a unified time; optimize variables , the variable time shift used in the solution; takes the value of a real number, and is the sole independent variable of the objective function; Robust loss A unified operator that employs Huber-type loss to suppress anomalous latency, combat outlier alignment errors, and applies saturation penalty to residuals is used. ; where scale parameter Defined by SLA and time granularity; output is a non-negative real number; regularized weights. : Sparse regularization coefficients, which are non-negative real numbers; anchor index : Index of the anchor set; The above estimation yields clock deviation Subsequently, all sentence segments are mapped to a unified timeline, and then multi-source texts are concatenated using this unified time as the horizontal axis, ensuring that subsequent turn-taking and evidence anchor point location share the same base time. The final output is a time-series index. By stably binding the (document identifier, sentence / segment position, and unified time) triple, cross-document playback and positioning can be supported.

[0029] In practice, alignment errors are suppressed by robust loss, and cross-timezone files and delayed emails are unified to the same reference coordinates; anchor-driven operation makes the unified timeline interpretable with business semantics; the output time-series index... It provides a reusable time base for subsequent event extraction.

[0030] Business conversations often involve a mix of long emails, short instant message replies, and meeting minutes summaries. Simply using line breaks or punctuation to segment conversations can lead to numerous misjudgments of turn boundaries. Therefore, constructing turn boundary energy is crucial. It integrates semantic embedding variations, pause duration, and punctuation signals into a single score, and searches for boundary indicator variables globally. The optimal configuration, combining sparsity and nearest neighbor constraints, yields a stable talk wheel structure, as follows: Where: Boundary energy Sentence / Paragraph The strength of evidence at the turn boundary is a non-negative real number; semantic embedding vector Sentence / Paragraph The contextual semantic representation, for , For dimensions; Corresponding sentence / paragraph ; Pause duration Sentence / Paragraph and The uniform time interval is a non-negative real number; punctuation strength Syntactic / punctuation evidence at sentence / segment boundaries, supplementing written signals, and are non-negative real numbers; weighting coefficients. : Multi-evidence weighting parameter, which is a non-negative real number.

[0031] After obtaining the local energy, the boundary set is obtained through global optimization with sparsity and nearest-neighbor penalty constraints. :

[0032] Where: boundary indicator variable Sentence / Paragraph Is it an indication of a boundary? Sparsity penalty coefficient : Controls the number of boundary values, which are non-negative real numbers; For sentence / segment index; Nearest neighbor penalty coefficient To suppress excessively dense boundaries, a distance-dependent second-order penalty is introduced, which is a non-negative real number; the decay factor... Distance attenuation parameter, is Boundary set : The set of selected boundary locations.

[0033] The boundary indicator optimization employs a two-stage method combining sequential dynamic programming and local coordinate descent. Stage one involves fixing... The DP global partitioning is performed based on the energy monotonicity; in stage two, closed-loop updates are used when the boundary structure is fixed. (Or grid search) to minimize the oversplitting rate of the validation set, up to 3 rounds.

[0034] Therefore, local energy unifies semantic mutations, prosodic breaks, and syntactic signals to a comparable scale, thereby enabling global optimization to generate robust boundaries under the tension between sparsity and proximity; furthermore, these boundaries are mapped onto a unified time axis as a turn sequence, thus forming a temporal index. The basic framework; the final A stable interface is exposed to subsequent steps using a unified time-document-segment three-way mapping.

[0035] In use, complex hybrid genres are segmented into replayable talk units; nearest neighbor penalty avoids jitter segmentation in short periods of time; timing indexing. It can directly support the location and playback of evidence anchor points.

[0036] Step 102: Based on a unified time and turn structure, group pronouns, ellipses, and cross-language aliases into stable references, and output a reference alignment mapping table. .

[0037] Business English documents contain job titles (e.g., Sales Lead, PM), institutional abbreviations (e.g., PO, ETA), and cross-language names and company aliases. Without standardization, participants and objects of events can be counted repeatedly or incorrectly segmented. Furthermore, pronouns and ellipses appear frequently in real-time communication; if they are not referenced back within the constraints of time and context, consistency among participants in event units will be disrupted. Therefore, it is necessary to standardize the mention set. With entity set The correspondence between them is solved in one go through optimization, and then pronouns and omissions are analyzed using a time-constrained relational model to form stable references and register evidence anchor indexes. .

[0038] To unify the aliases and spellings from different sources, the reference set will be... Viewed as supply distribution, the entity set Treating it as demand distribution, through the cost matrix The algorithm measures the combined cost of contextual semantics, spelling differences, and temporal consistency, and introduces entropy regularization and temporal priors to obtain smooth and interpretable alignment, outputting the optimal transfer matrix. :

[0039] Where: transmission matrix The matching strength of entities is mentioned as a non-negative matrix that satisfies edge constraints; edge distribution. The quality distribution of the entities mentioned is non-negative and uniformly normalized; cost matrix. The matching difficulty is measured by a non-negative real number matrix, taking into account semantics, spelling, and time cost; entropy temperature. Smoothing coefficient, a positive real number; time series prior matrix Prior matching from unified time and role constraints, a non-negative matrix; prior weights. : Controlling prior influence, compromising data cost with prior knowledge, non-negative real numbers; feasible set :satisfy A set of.

[0040] Therefore, optimal transmission produces sparse and stable matching under both semantic and temporal constraints, thereby unifying cross-document aliases to entity sets. The stability reference; and then the optimal transfer matrix is ​​determined by threshold and connectivity rules. Merge into reference alignment mapping table This allows for the direct use of stable reference-evidence anchor pairs in subsequent event units. Ultimately, Each mapping in the document is accompanied by an evidence anchor index. The source location is ensured to guarantee auditability. During use, cross-language names, abbreviations, and aliases are unified into stable references; matching is smooth and non-divergent, avoiding mismatches caused by local noise; the generated reference alignment mapping table... This directly serves the standardization of participants involved in subsequent event extraction.

[0041] For pronouns such as "he," "they," and "that one" in real-time communication, under the constraints of unified time and turn-taking structure, an interpretable log-probability model is used to synthesize semantic similarity, role compatibility, temporal proximity, and process consistency to obtain the posterior of anaphora:

[0042] Where: posterior probability of retracement Mention Pointing to entity The probability, Semantic similarity Mentioning semantic similarity to the entity's context, measuring semantic matching, usually normalized to 1. or Character compatibility Based on the match between job title / responsibility role, Recent score Based on a uniform time interval, the distance is a monotonically decreasing function, and is a non-negative real number; the probability of process consistency. Based on the reachability of the process state machine (e.g., submit PO → issue invoice → ship). Weight : The intensity of the influence of each factor, a real number.

[0043] Therefore, the anaphora of pronouns / ellipsis converges to a unique and interpretable entity under time and process constraints; thus, the high-confidence anaphora increment is written into the referential alignment mapping table. And register the corresponding original text location in the evidence anchor index. Furthermore, the backreference enhances and feeds back into the prior knowledge of optimal transmission. This forms a collaborative closed loop of alias regularization and anaphora resolution; ultimately, stable referencing permeates all turn-of-sight, ensuring unambiguity among event participants across documents. Pronouns / ellipsis maintain stable anaphora under the dual constraints of temporal and procedural consistency; multi-factor weighting provides an interpretable parameter tuning interface; and referential results serve as anchors of evidence. Synchronous output meets auditability requirements.

[0044] Normalize cross-source discourse segments into a set of event units containing trigger words, participants, objects, and time conditions. Furthermore, across document scales, event units pointing to the same real change are grouped into event clusters driven by auditable energy and prior knowledge. By explicitly embedding a uniform time, stable reference, and evidence anchor index during the extraction process. Then, in the merging process, consistency maps and entropy-temperature aggregation are used to achieve the sedimentation of factual references.

[0045] Step 201: In the document time-series index Alignment Map with Reference Under the common constraints, the sentence segments are precisely organized into event units of trigger word-participant-object-time condition, and evidence anchor indexes are bound to each field. .

[0046] Business English texts exhibit both cross-genre language phenomena and cross-channel record differences. Relying solely on sequence labeling can easily lead to issues such as trigger word drift, repeated participant counts, and mismatched time phrases. Because... Segments from different sources have been mapped to a unified timeline. Since pronouns, omissions, and aliases have been unified into stable references, a structured objective with both hard and soft constraints on temporal consistency and reference consistency should be introduced in the extraction stage. At the same time, temporal conditions must be calculably anchored to specific points on a unified timeline to ensure subsequent causal arrangement and evidence playback.

[0047] When a unified time and stable reference are available, the extraction problem is rewritten as a structured optimization with consistency regularization. Indicator variables are constructed from a set of candidate segments, and conditions such as unique triggers within the same talk turn, triggers and participants appearing within an reachable time window, and participants and objects coming from stable references are encoded as violation variables, thus obtaining the decision vector. :

[0048] Where: event indicator vector : A binary vector indicating whether a candidate fragment constitutes an event unit. ; Candidate features , No. The input features of each candidate (sentence segment or fragment) include lexical, dependency, reference mapping, temporal anchor and turn context; the values ​​are real vectors. Feasible set The set that satisfies the basic syntax and turn-taking mutual exclusion constraints is the discrete feasible region; Weight vector : Model parameters used for linear scoring, which are ; Feature mapping Multimodal features of candidate segments encode semantics, syntax, and position into the same space, such as linear sparse mapping, random Fourier mapping (kernel approximation), etc. Pair Consistency Weight Candidate The compatibility / mutual exclusion strength; the consistency or repulsion strength between candidate pairs, a real matrix; , Instructing candidates Whether it is activated; the value can be 0 or 1. Constraint Set : A predefined set of business consistency constraint indexes, which enumerates all violations that need to be penalized and is a finite set of indexes; Penalty weight : The weight of each violation, a non-negative real number; Violation quantity :exist and The degree of violation under constraints, measuring issues such as time window out-of-bounds errors and inconsistencies in attribution, is a non-negative real number. It is a set of pluggable non-negative scales used to measure the decoding result of the current event unit. The degree of deviation from various business / discourse rules; among which For document time-series indexing (unified timeline). This refers to the alignment mapping table (stable reference library). Each constraint... Each corresponds to a computable line. A value of 0 indicates compliance, while a positive value indicates violation; a larger value indicates a greater deviation. (This refers to all...) By weight Aggregates into the objective function, corresponding to candidate selection variables This leads to a solution that satisfies both the template and the uniform time and stability indices; weights Given by the validation set or business rules.

[0049] Directly achievable constraint types include: time window consistency (the deviation between candidate time and unified anchor point does not exceed the set tolerance); stable reference consistency for roles / objects (participants and objects must be able to find a credible mapping in the stable reference library); unique triggering within the same turn (at most one trigger word is activated in the same turn); unique field occupancy and template completeness (all fields required for a trigger must be present, and the capacity of each field is limited); process reachability consistency (the combination between triggers and fields must be consistent with the process prior or automaton reachability); near-duplicate suppression across documents (candidates that are close in time and highly similar in semantics should not be activated simultaneously); anchor point coverage (selected candidates must have a replayable position in the evidence anchor point index); and relative order consistency (candidates with sequential relationships must not be reversed on the unified timeline). The input of these quantities depends only on the preceding S1 product. The algorithm calculates the original text position, timestamp, and semantic similarity of the candidate text with the candidate text, and outputs a non-negative score per constraint. In implementation, these items are calculated and summed module by module; during training or parameter tuning, only adjustments are needed. With tolerance threshold. The decoding result will write the activated candidate along with the original text offset, source document and unified timestamp back to the evidence anchor index. The subsequent step three is used for chain and graph construction, step four is used for path constraint compilation, and S5 is used for auditable return, forming an end-to-end consistent coordinate system and rule caliber.

[0050] In practice, this objective unifies trigger word localization, role binding, and object normalization within the same framework, thereby reducing cross-field error propagation; the consistency second-order term stabilizes the unique triggers within the same conversation round and the mutual exclusion relationship of adjacent segments; violation pairs and The explicit dependency ensures a strong binding of the output event unit to uniform time and stable reference.

[0051] Business English often uses relative time expressions such as "by EOD Friday," "next quarter," and "within 48 hours." It is necessary to find a single or interval anchor point on a unified timeline and establish a revisitable connection with the original text.

[0052] The candidate time anchors are matched with the relative phase of the time expression using a phase alignment kernel, and the solution is maximized using the compatibility provided by the turn-by-turn context as weights.

[0053] Where: Event time estimation The unified time anchor point for event units; candidate set From time-series index Extracted event units Relevant time candidates, limiting the search domain; Anchor Time From A unified time anchor point (using the semantics of step one) provides an alignment reference; frequency parameters : Frequency coefficient of the phase kernel, a positive real number; optimization variable : The time position to be searched, with a real number as the value; compatibility score : The degree of fit between time expressions and candidate anchors in turn-taking and syntax, a non-negative real number.

[0054] In use, phase-aligned kernels are inherently robust to periodicity and relative expression, allowing mapping of weekdays / month-ends / quarters to specific anchor points or interval centers; and The candidate constraints work together to avoid cross-document temporal drift; the output event time estimate Registered simultaneously with the original phrase location in the evidence anchor index. This provides a direct entry point for subsequent auditable playback.

[0055] Step 202: Set up the event unit collection Event units that refer to the same real change cluster together into event clusters across document scales. Each event cluster is assigned a unique factual reference and an interpretable consistency score.

[0056] The same business fact is often repeatedly described in multiple source texts at different granularities. For example, emails emphasize that the Purchase Owner (PO) has submitted the document, instant messages emphasize approval and release, and meeting minutes record that the goods will be shipped next week. If cross-document merging is not performed, redundant nodes will be formed in the process graph, and the sparse structure of the causal chain will be destroyed. Due to Stable designations for participants and objects have been provided. A unified time anchor point has been provided. Since evidence has been provided, the merging problem can be characterized as a collaborative optimization of prior-similarity-connectivity, which must respect the hard boundaries of business roles and time accessibility, while maintaining the smoothness and interpretability of clustering under the diversity of expressions.

[0057] A consistency graph is constructed for event units, with edge weights reflecting trigger word semantics, stable referential consistency, and temporal reachability. To avoid early errors caused by greedy merging, soft allocation agglomeration is performed using the entropy regularization objective of the row random allocation matrix, resulting in a soft allocation matrix. :

[0058] Where: soft allocation matrix The allocation strength from event units to event clusters is a row-random nonnegative matrix; feasible region : A set whose sum is 1 for each row ; For the index of the enumerated event cluster; adjacency pair set For index pairs of edges taken on the event unit adjacency graph, only neighbors that are temporally proximate, share the same turn, share participants / objects, or have the same trigger word family are included; soft weighting is applied. Used to quantify matching strength; prior weights : A priori positive real matrix consisting of stable referential consistency, triggering semantic similarity, and temporal proximity; Consistent edge weight : Connectivity strength of event unit pairs, a non-negative real number; row vector : No. and Row allocation vector, for ; When used, the first objective transforms the merging problem into minimizing relative entropy, ensuring that the assignment follows the prior but is not overly sharp; the second objective smooths the assignment on the graph structure, avoiding misclassification of events with diverse expressions but consistent facts; soft assignment yields hard clusters after subsequent thresholding or temperature annealing, and the process preserves the confidence trajectory of each event and each cluster, facilitating auditing.

[0059] After soft allocation is completed, a unique and reusable fact label needs to be generated for each event cluster. This label is then used to inversely constrain the event units whose values ​​are still uncertain, forming a closed loop of extraction-merging-anti-disambiguation. The cluster center is then obtained in the embedding space using the Bregman mean, and fine-tuned using business invariants.

[0060] Where: cluster center vector Event cluster Factual reference embedding; optimization variables The variable center during the solution process; Event Embedding Vector Semantic-role-temporal composite embedding of event units provides a metric basis; Bregman divergence. : By differentiable convex functions The derived divergence characterizes the center under non-Euclidean geometry; Invariant vectors : Anchored factual references based on priors extracted from stable references and business rules (such as fixed material codes or contract numbers); invariant weights : To balance prior knowledge with data, resulting in a non-negative real number.

[0061] To ensure interpretable consistency metrics and refine low-confidence boundary samples through backfeeding, a consistency score for event clusters is defined:

[0062] Where: Consistency score :cluster Overall coordination Pairing consistency :based on Time available The combined score of participant consistency and trigger similarity measures the reasonableness of co-clustering between two event units, providing a basis for... Soft weight allocation As defined above, weighted average of pairwise consistency; .

[0063] In practice, the Bregman center ensures that factual claims have minimal divergence in non-Euclidean space, accommodating asymmetric similarity and long-tailed representations; consistency scoring provides scalar signals for auditing and automatic annealing, and back-features low-scoring samples to step 201 to strengthen the re-evaluation of trigger or time fields; factual claims With consistency score Registered together in the evidence anchor index This ensures that the merger decision can be reviewed.

[0064] With event clusters Cluster center vector Consistency score Input the structure and index the document time sequence. Alignment Map with Reference Under the dual constraints, the event chain construction and process graph generation are completed. In a unified time and stable nominal coordinate system, event clusters are arranged into event chain sets according to their sequence and reachability. This leads to the generation of process maps. , where the node set Instantiated from event clusters, relation edge sets Simultaneously, causal relationships and state changes are marked, and evidence anchor indexes are attached to each relation edge. Auditable references are used to ensure that the sorting is verifiable, the source is traceable, and the status is explainable.

[0065] Step 301: In the event cluster set Cluster center vector Consistency score On the input, combined with document time-series index A unified time and business reachable priors are obtained, leading to an acyclic global order that is segmented into a set of replayable event chains. .

[0066] Business facts often involve multiple sources and varying granularities. Directly sorting by single-point time is susceptible to local delays and redundant expressions, leading to chain breaks or pseudo-concurrency. Given that step 202 has already output cross-document aggregated event clusters and factual claims, and the document temporal index... A unified time anchor point is provided, requiring the construction of global order variables to absorb local noise. Simultaneously, compliant processes (such as order placement → invoicing → delivery → payment collection) must be transformed into executable hard constraints, ensuring that the ordering conforms to temporal evidence without compromising process accessibility. Ultimately, local adjacent segments are identified within this order, forming multiple event chains, thus providing stable candidate sequences for edge generation in the process graph.

[0067] Define the chain order rank for each event cluster Using paired interval loss to drive the established sequential relationships and to set the hard constraints of the process. Adding a unidirectional, irreversible order constraint to the optimization yields:

[0068] Where: chain order rank , Event cluster , Ordered embedding on the global timeline; Weight matrix elements Cluster pairs The strength of evidence in sequence, by Time difference, consistency score With consistency score The confidence level and the consistency of stable indices are jointly generated; Interval loss : Promoting chain order rank At least leading chain rank A single unit interval avoids sequence sticking caused by weak evidence; a set of hard constraints on the process. : A set of necessary sequential relationships extracted from business accessibility and compliance rules to ensure that the order does not violate the process; order penalty coefficient Control the priority of hard constraints to prevent time noise from overriding the inevitability of the process; In practice, the global order variable absorbs local temporal noise at intervals, ensuring a consistent order even in multi-source asynchronous environments; hard process constraints transform business priors into irreversible order constraints, thus avoiding solutions that violate the order of processes; optimization results can be directly applied to... Top alignment facilitates evidence playback.

[0069] After obtaining the order, it is necessary to determine which adjacent or nearest neighbor pairs constitute causality, and simultaneously label the transition categories of object states. The causal posterior of edges is modeled using an interpretable log-odds model, and process reachability is used for gating, resulting in: Where: causal posterior probability : Directed pair The probability of causality exists is used as the main score for edge selection; Weight vector : The strength of the influence of features, and the weight of learning time and business factors; Feature vector : By chain order difference Precursor matching of object roles, cluster center vector with cluster center vector The state difference mapping and the consistency of evidence constitute the basis for causal determination; it is a real number vector.

[0070] Therefore, candidate edges are formed using a threshold. and with process reachable gating instructions Illegal transitions are filtered out to obtain a set of edges that are causally consistent and reachable. Then, a state decoder is used to process the cluster center vectors. Mapped to a state vector, and with the cluster center vector The mapping difference is used to generate discrete transition labels for creation, approval, performance, and settlement, and then adjacent high-confidence edges are merged into event chain segments and summarized as follows. .

[0071] As a supplement, the state decoder is defined. Using gated projection + simplex normalized output for creation / approval / fulfillment / settlement / anomaly handling, etc. Class probability vector, state label taken . Supervised by labeled small samples, with the remainder expanded by constrained self-training; first based on... and In the time direction, construct the maximum acyclic subgraph (Kahn topology edge deletion to minimize the total inverse hierarchical loss); then perform hierarchical optimization, and finally perform evidence sparsification on the edges (retaining the first edge value for each edge). The highest (Anchor point entry into the warehouse) When used, causal determination is data-driven and process constraints are controllable, avoiding false edges that are adjacent in time but unrelated to business; state labels and causal edges are generated simultaneously, enabling chain segments to carry semantic transition interpretations; chain segment division is completed in the order space, reducing interference from differences in cross-document expression.

[0072] Step 302: Set the event chain Projection as a process map It outputs edges with causal and state labels using a hierarchical layout and evidence weighting, while also indexing the evidence anchor points. Perform auditable binding.

[0073] The process graph serves as both the operational vehicle for subsequent retrieval and reasoning, and the structural foundation for the interconnected presentation of paths and evidence. Without hierarchical organization and evidence allocation, the graph will lose readability due to crowded edges and will struggle to return actionable evidence packages during queries. Based on the causal candidate edges and chain segments obtained in step 301, clear hierarchical structures need to be assigned while ensuring directed acyclicity; simultaneously, each edge must be bound to an anchor weight pointing to the original text, enabling the answer to be reviewed and the path to be replayed.

[0074] To improve readability and query efficiency, hierarchical variables are assigned to each event cluster node. It also penalizes inverse hierarchy and insufficient relaxation on the candidate edge set, while incorporating the hierarchical priors given by the business. ,get:

[0075] Where: hierarchical variables , :node , The graph hierarchy determines the visualization and traversal order; edge sets : The set of causal reachable edges selected in step 301, which serves as the carrier of hierarchical constraints; Inverse hierarchical loss This ensures that upstream levels are lower than downstream levels and that there are boundaries, avoiding overlap at the edges; Hierarchical Priors Based on the prior knowledge of the business process (such as procurement to payment), anchor the node level; The mapping table from the process automaton stage ID to the node status label is used to determine the causal strength when conflicts occur. The topological sorting result is the primary factor, with prior information serving as a secondary penalty; the hierarchical penalty coefficient... : To balance prior knowledge with data and prevent deviation from common business sense; it is a non-negative real number.

[0076] When used, hierarchical optimization maximizes the downward flow of edges while maintaining directed acyclicity, greatly reducing visual and traversal conflicts; prior terms inject domain knowledge into the layout, making the graph both interpretable and stable; hierarchical variables provide natural path constraints for query compilation.

[0077] Each causal edge requires organizing the original text's anchor points into a traceable evidence package. To this end, the edges... The semantic key vectors are bilinearly matched with the text vectors of candidate anchors, and evidence allocation is formed using normalized weights:

[0078] Where: weight of evidence Anchor point opposite side The strength of belonging is And for The sum is 1; Edge semantic key vector :Depend on and Obtained through edge transformation, characterizing edge semantics; summation index. Iterate through the denominator. The anchor index is used to perform Softmax normalization; the anchor text vector From the anchor index of evidence Embedding of segments, measuring the matching degree with edges; edge evidence set. : By nodes The relevant original text anchor points are merged and deduplicated to obtain the allocation domain.

[0079] Therefore, evidence allocation establishes a one-to-many interpretable link between the edge and the original text fragment; then, it writes the highest-weighted anchor points, their source documents, and unified timestamps into the edge attributes, ultimately forming a three-in-one edge object of causality, state, and evidence; and further retains this information in the graph storage layer. Weighted trajectories ensure that audits and reviews can be replayed.

[0080] When used, evidence weighting avoids black-box responses that only provide an answer without evidence; edge objects carry lightweight but sufficient evidence indexes in storage, enabling subsequent queries to aggregate evidence fragments in constant time; the aggregation process does not destroy... A unified time coordinate facilitates multi-source playback.

[0081] In process graph With event chain collection Assuming stability has been achieved, the conversion of natural language to structured graph paths for real-world business retrieval is engineered: first, the user query... Parsed into a query intent expression And compile the roles, objects, and time constraints into a set of path constraints. To make it compatible with time-series indexes Alignment Map with Reference Maintain the same coordinate system; then generate a joint text and graphic execution plan. In relational edge set The above is based on causal posterior probability. Constrained path search, while indexing from evidence anchor points Mid-scale extraction of reviewable evidence packages, and weighting of evidence. The data is collected to form an auditable plan with a dual path and evidence channel that can be directly consumed in the execution phase.

[0082] Step 401: Transfer user query Compile into a query intent expression in a coordinate system with unified time and stable reference. With path constraint set This translates the natural language conditions of role-object-time into executable constraints on the graph.

[0083] Business English searches often combine task objectives (such as confirming whether PO approval has been completed and providing evidence) with background context (such as after a customer's payment reminder call last week).

[0084] If not using process diagrams Structured compilation for prior knowledge degenerates retrieval into plain text matching, failing to guarantee consistency between the returned path and the evidence. This is due to the event cluster set. It has been settled into fact nodes and has cluster center vector event chain set The reachability order and causal posterior probability are given. The edges are given a credibility scale. Therefore, incorporating these existing quantities into semantic parsing and constraint generation makes the language-to-path mapping interpretable, revisitable, and reusable in the execution phase.

[0085] In situations where natural language phrases are interwoven with business terminology, to avoid abstract semantics deviating from the process map... By incorporating textual evidence, stable references, and chained priors into the same optimization framework through energy-based joint decoding, query intent expressions are directly generated. (Includes target node, constraint predicates, and output format):

[0086] Where: Query intent expression The semantic parsing output, including the target, filtering predicates, and output template, serves as the entry point for subsequent path compilation; candidate set. Allowed intent expression space, restricting invalid structures; weight vector Local text scoring weights measure the matching of word fragments with intent slots; Feature mapping User query The local features provide linguistic and contextual information. The vectors extracted from the query text can be composed of pre-trained semantics, keywords, regular phrases, temporal expressions, and channel priors, such as lexical template mapping (suitable for operators that select start / end / necessary links), semantic attention mapping (aligning operator cues with pre-trained encoders), and time window mapping (used to set / parse time range operators). Constraint Index Set : A set of business consistency and coordinate consistency constraints, enumerating violations, which is a finite set of indexes; penalty weights. The weight of the violation term is a non-negative real number; Violation quantity Relative time-series index Refers to the alignment mapping table With event chain collection The consistency violation degree is used to analyze the contradiction between the suppression and the existing structure; a value of 0 indicates satisfaction, while a value of positive indicates violation. Node set : A set of process graph nodes, carrying factual claims; prior weights : Graph prior strength, moderating the coupling between semantics and structure; is a non-negative real number. Matching indicator Intended to assign to a node The indicator quantity activates the graph prior; for .

[0087] Compatibility rating Cluster center vector With query The compatibility of the data allows the semantics of facts to be injected into the parsing, resulting in real numbers.

[0088] As a supplementary explanation, achieving consistency means... The specified start point, end point, necessary and prohibited conditions are mapped to The process involves finding a set of candidate paths that meet the given conditions and calculating the insufficient quantity based on the reachability strength of the edges; time window consistency will... Projecting the time range onto a unified timeline This measures whether the time interval between adjacent events within a candidate path exceeds the tolerance and whether the overall path falls within a given time window; role / object consistency... The characters and objects mentioned in the text are through Mapping to stable references, checking if a high-confidence match can be found in the participant and object fields of the path nodes; prohibition stage constraint statistically analyzes whether nodes marked as prohibited are traversed on all candidate paths; node anchor confidence separation requirement. Key phrases need to be clearly anchored to a small number of nodes, and the semantic similarity difference between the preferred and second-best nodes must exceed a set interval; path complexity / length constraints measure the shortest path that meets the conditions, and a violation occurs if the number of edges exceeds the upper limit. All of the above rely only on the defined... , , The output is a non-negative number for each constraint, corresponding to the query text. As a punishment, weight The results are derived from business configuration or validation set calibration, ensuring that the parsing results are consistent with the unified time, stable reference, and process graph in the same coordinate system.

[0089] Matching Indicator This indicator is used in the parsing and scoring in step S4, and its value is only 0 or 1, representing the query intent expression. Whether the process graph nodes are clearly reached The system calculates using a two-path parallel approach, taking the maximum value: the first path is the anchor hit, from... Extract the anchor phrase set from the middle, and use the node cluster center vector Compatibility score with phrase semantic vectors Matching is performed when the score is not lower than a set threshold. Time node The first path is recorded as a hit; the second path is a path inclusion, based on... The starting point, ending point, necessary and forbidden clauses, in the process graph Retrieve the set of all candidate paths that satisfy the constraints, as long as there exists a path containing the node. This is considered as reaching the target. Either path being completed immediately triggers an order. Otherwise, it is 0. This indicator does not depend on continuous hyperparameters outside the threshold and is consistent with the unified time index. Alignment Mapping Table Maintaining the same coordinate system: path retrieval dependency The reachability structure, anchoring the cluster center vector of the hit dependent node and the semantic vector of the query phrase, and the hit determination threshold. It can be configured by development set quantiles or fixed values, and maintains a stable caliber across different business domains. The results are used both for structured scoring in the analysis phase and as a priori selection switch for subsequent evidence presentation sub-plans.

[0090] Compatibility rating This score is a non-negative semantic fit metric, measuring the query text. With nodes The semantic consistency is used as an alignment reward in the parsing objective of step S4 and as an importance prior in evidence presentation. The calculation process is as follows: First, the nodes are... Cluster center vector semantic vectors of query text Normalize each vector to obtain vector pairs that directly correspond to the cosine similarity; then apply temperature scaling (temperature parameter) to the inner product of the two vector pairs. (Controlling the smoothness and separability of the fractions), then mapping them to the non-negative real number domain via a softplus transformation to avoid instability caused by negative fractions and maintain the same dimensions as other summation terms. The resulting... It's not a probability, but a comparable compatibility strength, comparable to a threshold. It can be used in conjunction with anchor hit determination and can also be directly sorted among multiple nodes. In implementation, Obtained from a pre-trained semantic encoder, it can be overlaid with domain adaptation; Validation set calibration can ensure stable distribution under different query lengths and node densities; when a query is split into multiple phrases, the distribution of each phrase is calculated separately. The maximum value or a weighted aggregation is then selected to balance keyword anchoring and sentence semantic alignment. Both are combined with... , , To maintain consistency in the unified approach, ensuring reusability throughout the entire process of parsing, path constraints, evidence selection, and auditable return.

[0091] Therefore, semantic parsing is constrained by four-way evidence—text, graph, temporal, and referential—to avoid misclassifying ambiguous pronouns or cross-linguistic aliases to the wrong objects; and the parsed output is then used to express the query intent. The form carries the target node and filter predicate, and is linked with the time-series index. Binding time domains and aligning them with reference mapping tables Binding role / object domain; ultimately a set of path constraints. The generation of these anchors establishes stable semantic anchors.

[0092] When using it, the analysis results are used for the process graph. It possesses native compatibility; violating quantitative and prior weighting ensures that parsing neither goes to extremes nor becomes overly ambiguous; the output query intent expression It can be directly consumed by subsequent compilers.

[0093] Based on query intent expression The slot content systematically translates the start / end / necessary / prohibited steps that users care about into a set of path constraints. To ensure that the constraints and causal posterior probabilities are consistent With Time Series Index Collaboration, defining path candidates Constraint consistency objective:

[0094] Where: Path candidates : In relational edge set Directed path on the spectrum, the execution skeleton being scored; causal posterior probability :side Causal confidence (using the aforementioned definition). Constraint elements : From the path constraint set Single constraints, characterizing must / prohibit / time window / role binding; penalty coefficient. Time consistency penalty strength, a non-negative real number, weight. Constraint importance, adjusting the impact of different constraints; satisfying instructions. :path Does the constraint satisfy? The instruction transforms logical constraints into points-based rewards. Timing gap penalty :path In time-series index The gap cost below, suppressing jumps inconsistent with the unified time, is a non-negative real number; Defined as the weighted sum of the overthreshold intervals of adjacent edges at a uniform time and the number of jumps across documents; the weights are adjusted by the playback experience annotations.

[0095] Therefore, the path constraint set It is explicitly mapped to an optimizable scoring function, which both strengthens the coherence of high-confidence causal edges and uses temporal gaps to suppress leapfrog splicing; then, according to the constraint consistency objective... To maximize the performance, several parallel candidate paths are retained. Finally, the candidate paths, constraint satisfaction, and time smoothness are output together to form the graph-text joint execution plan. It provides a quantitative interface for scheduling and evidence orchestration.

[0096] When used, path scoring combines semantic goals and structural evidence into the same optimizable dimension; constraint satisfaction terms make mandatory / prohibited / time windows / roles measurable; and temporal penalties ensure that the path is revisitable and interpretable.

[0097] Step 402, in the path constraint set Generating a joint graphic and text execution plan under the drive At the same time, indexed by evidence anchor points Weight of evidence Organize an evidence package within the budget to ensure that subsequent execution phases can return an auditable result of answer-path-evidence in one go.

[0098] The value of business search lies not only in finding nodes or paths, but also in delivering actionable evidence packages. Evidence Anchor Index While the evidence has been stabilized and accumulated in previous steps, without budgetary selection and redundancy suppression, it will appear as a jumble of noise, affecting readability and auditing. On the other hand, the feasibility of the path itself needs to consider causal posterior probabilities. Multiplicative effects and time series indexes Smoothing constraints. For a given path candidate First, the edge-level evidence set is aggregated on the set of edges involved in the path to form a path evidence candidate domain. In order to be within the budget To balance coverage and weighting, this scheme uses a joint modeling approach with kernel determinant and weighted linear terms:

[0099] Where: Evidence selection set The final set of anchor points for evidence constitutes the evidence package. A finite subset; Path evidence candidate domain : By path The set of evidence involved is the union of the sets of evidence, which limits the selection range; kernel matrix :exist The similarity kernel defined above, with To encourage diversity, a symmetric positive definite matrix is ​​used, and a convex combination of text embedding cosine kernel and time nearest neighbor kernel is adopted to ensure that semantic redundancy is suppressed and evidence crowding at the same time is avoided. submatrix : For sets The core matrix measures the coverage and diversity of the selection set; it is a symmetric positive definite matrix; the weights... Anchor of evidence Importance score, combined with evidence weight The value obtained from the source credibility is a non-negative real number; the tradeoff coefficient : Adjusts the relative strength of diversity and importance, adapting to different business preferences; is a non-negative real number; budget : The upper limit of evidence items, controlling the presentation scale and reading cost; it is a positive integer, automatically issued based on the query intent template and business SLA (such as ≤10 pieces of evidence to be issued by the audit), and different upper limits are set for mobile and desktop terminals respectively; therefore, This suppresses crowding of homogeneous evidence, thus covering more aspects of the facts with a few anchor points; while the weighting term assigns weight to the evidence. The explicit inclusion of authoritative source documents prioritizes high-confidence, high-diversity combinations, thereby selecting the final set. Index by Time Sorting to generate evidence presentation sub-plan And bind each piece of evidence to the position of the corresponding relation edge on the graph one by one.

[0100] In use, the evidence package is controllable in size yet high in information density; similar evidence is suppressed while key evidence is preserved; and the output evidence presentation sub-plan... With path Native alignment makes it easy to consume directly during the execution phase.

[0101] To unify path reliability, time smoothing, and evidence credibility onto the same quality scale, graph path selection and evidence selection are merged into a plan. The quality of the plan is defined by a differentiable weighted logarithmic objective:

[0102] In the formula: plan : By path With evidence set The joint plan constitutes the direct input during the execution phase; causal posterior probability. :side causal confidence, for gap set :path The set of time intervals, based on the time sequence index. Definition, measurement of continuity gaps; gap cost :gap The cost, which increases with the gap length and cross-document jumps, is a non-negative real number; the penalty coefficient... : Time smoothing penalty strength, a non-negative real number; Evidence weight Anchor of evidence Relative to its belonging edge The weight is Trade-off coefficient : Evidence contribution strength, balancing path confidence and evidence confidence; is a non-negative real number.

[0103] Therefore, the generation of a joint text and image execution plan is reduced to a matter of plan quality. Maximize; first, prioritize path consistency. Select several high-scoring paths, and then obtain the evidence objective for each path. Finally Perform reordering and scheduling to output one or more parallel executable plans; then, present the evidence in the order of presentation, timestamps, source document identifiers, and evidence anchor indexes for each edge in the plan. The two-way association ultimately forms an integrated instruction consisting of an answer paragraph template, a path node sequence, and an evidence package list.

[0104] When used, the constructed quality function combines path reliability with evidence credibility, which can avoid biased outputs that only have paths without evidence or only have evidence without context; the scheduling results can be directly replayed in a unified time and stable reference coordinate system, meeting the needs of auditing and review.

[0105] Joint Implementation Plan for Graphics and Text The implementation and auditable delivery, in the process graph The above is the query intent expression With path constraint set As an entry point, using causal posterior probability Weight of evidence To measure this, the optimal path is selected with risk aversion, and the evidence anchor index is aggregated. Generate a set of answer nodes With replayable answer-path-evidence results; simultaneously, constructing a chain-like audit trajectory. Time series index Alignment Map with Reference Coordinate constraints are applied throughout the record to ensure that the returned data is verifiable, traceable, and reproducible.

[0106] Step 501: In the joint execution plan of text and images On the candidate set, a joint selection based on risk aversion, combining path reliability and evidence credibility, is performed to obtain the optimal plan. and the corresponding set of answer nodes The results are then aligned in a unified time and stable reference coordinate system for review.

[0107] Intuitively choosing high-scoring paths greedily often amplifies risk due to the accidental high confidence of a single edge; while pursuing only the quantity of evidence sacrifices temporal continuity, leading to disjointed replays. Given that the graph layer already provides the causal posterior probability... Weight of side evidence With time gap cost Furthermore, the previous stage provided the mass function. By introducing entropy-temperature risk aggregation to absorb unilateral uncertainty, an interpretable compromise is formed among path reliability, time smoothness, and credible evidence, ultimately selecting the optimal execution plan that can be reviewed by the business.

[0108] Within the candidate path set, the peak value of unilateral uncertainty tends to dominate the overall decision. To address this, entropy-temperature aggregation is introduced, which smoothly maximizes the edge-level uncertainty at the path level after temperature adjustment, and then optimizes it in conjunction with the plan quality, thereby suppressing the risk of isolated peaks.

[0109] The path risk aggregation function and optimal plan selection are defined as follows:

[0110]

[0111] Where: Path risk aggregation : Measuring the path The entropy-temperature polymerization amount of accumulated uncertainty is smoothed with an adjustable temperature for the worst-case scenario; temperature parameter Adjusting the sensitivity to the maximum uncertainty to achieve a continuous transition from the mean to the maximum value; edge uncertainty It is a combination of the negative logarithmic causal posterior of the edge and the evidence dispersion, which quantifies the edge-level risk; Composition: as specified ,in Shannon entropy assigned to evidence (the more dispersed, the more uncertain). Learning from the validation set's risk preference annotations; optimal plan The selected combined text and image execution plan serves as the sole entry point for execution and return; quality function. The combined quality of path reliability, time smoothing, and evidence credibility (using the aforementioned definitions) is the main scoring criterion for the plan; risk weights. The weight of path risk in the overall objective is a non-negative real number; path Collection of evidence :plan The two parts respectively carry the structure and the evidence.

[0112] When using it, path risk aggregation is employed. Entropy-temperature aggregation smoothly amplifies unilateral risks, and extremely uncertain edges are no longer masked by the average; with the mass function The synthesis objective ensures a balance between high confidence, low gap, and strong evidence; outputting the optimal plan. It possesses robustness and explainability.

[0113] During execution, user questions may correspond to a single endpoint or a set of answers. To avoid directly equating path scores with answer confidence, edge-evidence-temporal joint normalization is performed on candidate endpoint nodes near the optimal path to obtain posterior confidence, and the answer set is determined under threshold and diversity constraints. The posterior of the answer node is defined as follows:

[0114] Where: posterior probability of the answer :node The posterior of the final answer; the primary metric for answer selection; Weighting coefficient : Measures the impact of causal strength, weight of evidence, and time gap, respectively; interpretable and learnable parameters; set of input edges. :path Inflow node The set of directed edges that converge to Causal evidence; lateral causal posterior :side causal confidence, reaching Structural strength, ; subset of evidence Belonging to a node The set of evidence anchors measures the textual support associated with each node; evidence weight. Anchor point Relative to its belonging edge Weighting, evidence credibility scale, Local gap cost :arrival The time gap before the jump is a penalty for a jump that cannot be replayed; it is a non-negative real number.

[0115] exist The next step is to perform a sub-modulus selection (greedy approximate maximization of coverage - distinct index), with Semantic similarity is used as a mutually exclusive kernel; this ensures that the answers are both highly confident and complementary.

[0116] In practice, the posterior answer unifies structural strength, evidence strength, and temporal continuity onto the same log-odds scale, avoiding skewed judgments that rely solely on any one dimension; threshold and diversity control enable multi-terminal problems to obtain a stable and complementary set of answers. ; and time-series index The binding guarantee ensures that every answer can be replayed in its original form.

[0117] Step 502: Develop the optimal plan. With the answer set The organization is a unified, replayable object of answer-path-evidence anchors, and the audit trail is constructed using chained hashing. To achieve compliant delivery that is non-repudiable and repeatable.

[0118] In business scenarios, traceability is emphasized: it's not just about presenting the conclusion, but also how it was obtained. Without presenting process metadata (such as query intent expressions)... Path constraint set Time-series index Refers to the alignment mapping table With Evidence Anchor Index The data is fixed upon return, which may lead to uncertainty due to secondary calculations during subsequent reviews.

[0119] To ensure non-repudiation, the trajectory records generated from each execution are linked into a hash chain in chronological order; any modification will disrupt the chain's consistency. Simultaneously, all evidence location information is uniformly indexed by time sequence. With Evidence Anchor Index At the coordinates, ensure cross-source replay. Chained hash recursion is defined as:

[0120] Where: hash value : No. A chain of fingerprints recorded from the traces, tamper-proof and verifiable; hash function. Collision-resistant hash mapping maps inputs of arbitrary length to outputs of fixed length; concatenation operation. Byte-level join, which concatenates the previous hash with the current record; Serialization function Transformation that converts trajectory record items into deterministic byte sequences eliminates implementation differences; By employing field sorting, unified units, and byte-level canonicalization (fixed UTC time, ISO document ID, and fixed number of digits for numerical values), we ensure consistency of the same object across platforms.

[0121] Track recording item : Contains query intent expression Path constraint set Optimal Plan Answer set Time-series index Refers to the alignment mapping table Evidence Anchor Index A tuple of timestamps and configuration summaries, the smallest unit of auditing; a structured object; initial hash. The fingerprint of the previous record is used to establish a chained association; this is related to the hash value. Same-domain elements.

[0122] In use, chained fingerprints ensure that any modification to any field will trigger a change in the end hash, achieving verifiable non-repudiation; unified With Evidence Anchor Index Coordinates ensure that cross-document evidence can be played back instantly; audit trail It can serve as the sole source of fact for subsequent model updates or dispute reviews.

[0123] To maintain reproducibility without disclosing sensitive information, minimal distortion anonymization is performed within the evidence text embedding space. A reproducible template is used to bind the path and evidence, ensuring that any third party with the same graph and coordinates can repeatedly verify the data. The anonymization optimization objective is defined as follows:

[0124] In the formula: desensitization embedding Anchor point The release vector approximates the semantics of the original text without revealing sensitive slots; Original embedding From the anchor index of evidence Text embedding; Bregman divergence : By convex functions Defined divergence, measuring the semantic deviation before and after desensitization; mask matrix. A mask that sets the sensitive dimension to 1, constraining the sparsity strength of the sensitive slots; ; The intersection of entity recognition, rule table, and outlier detection is used: the named entity and contract / number rule hit bits, and rare feature bits that appear only in a single document across the corpus are all set to 1. These 1-bit bits are then used as an irreversible hash field to retain auditing capabilities. Candidate denoising vectors. , used to represent the anchor vector of evidence An alternative, regularized representation; element-wise product Hadamard product, which restricts the sparsity penalty to the sensitive dimension; sparsity coefficients. : Controlling the desensitization intensity, a trade-off between semantic fidelity and privacy protection, is a non-negative real number.

[0125] When using, apply to the sensitive dimension. Sparsity allows for the controlled compression of key slots, while Bregman distance maintains semantic approximation of non-sensitive dimensions, thus enabling readable, verifiable, and non-leaking evidence release. Combined with a fixed template, the answer paragraph, path node sequence, and evidence list are output as a unified whole, allowing third parties to access the same information simultaneously. , , and The above can be used for re-verification.

[0126] Ensure that the returned data can be used directly, meet compliance and review requirements, and be reused as an audit-tagged data source for subsequent supervision and learning.

[0127] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0128] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0129] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0130] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0131] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for generating and intelligently retrieving knowledge graphs for business English scenarios, characterized by: include, Perform turn-taking and time-sorting on emails, meeting minutes, and instant communications, generate document time-series indexes, simultaneously produce a referential alignment mapping table, unify pronouns, omissions, and aliases into stable references, and register evidence anchor point indexes; Under the constraints of the document temporal index and the reference alignment mapping table, the sentence segments are normalized into event units containing trigger words, participants, objects and time conditions, and event units pointing to the same change are merged across documents to form event clusters; The event clusters are sequentially arranged according to the document time sequence index and the causal relationships and state changes are marked. The sorting results are mapped to event nodes and relationship edges of the process graph, and the evidence anchor point index corresponding to the original text position is recorded for each relationship edge. The user input is parsed into a query intent expression, and the roles, objects, and time requirements are translated into path constraints of the process graph. The evidence presentation sub-plan is compiled according to the evidence requirements, and the evidence anchor index is used to generate a set for consumption in the execution phase. Using the query intent expression as the entry point, the answer that satisfies the path constraint is located in the process graph. The corresponding fragments of the evidence anchor point index are aggregated, and the answer, path and evidence anchor point are returned as a whole. The structured trajectory of the query and return is saved.

2. The business English scenario knowledge graph generation and intelligent retrieval method as described in claim 1, characterized in that: The talk-turn segmentation and time sorting include robust time alignment driven by cross-source anchors to estimate clock skew in each document and establish a unified timeline; Global optimization is performed using the boundary energy derived from semantic embedding changes, pause duration, and punctuation intensity fusion. Dynamic programming and coordinate descent are used to solve the problem together, generating a set of turn boundaries and their unified timestamps, document identifiers, and paragraph offsets, which are then written into the document's temporal index.

3. The business English scenario knowledge graph generation and intelligent retrieval method as described in claim 2, characterized in that: The reference alignment mapping table is generated collaboratively by alias normalization and anaphora parsing. Alias ​​normalization adopts optimal transmission with entropy regularization and introduces temporal and role priors. Anaphora parsing combines semantic similarity, role compatibility, temporal proximity and process reachability. The role priors are derived from the process automaton and stable references. The parsing searches for candidate entities within a unified time window. Mentions that do not reach the threshold are marked for disambiguation and registered together with the original character range and source document in the evidence anchor index.

4. The business English scenario knowledge graph generation and intelligent retrieval method as described in claim 3, characterized in that: The event unit is generated using constrained structured decoding, including unique triggering of the same turn, participants and objects from violations, including the reference alignment mapping table, and time conditions constrained by the document temporal index. The violations are quantized positively and participate in the unified solution. The time condition parsing maps absolute dates, relative periods, and corporate terminology to a unified time anchor, records the original character offset of the quadruple field and the source document, and associates the evidence anchor index to form an event unit object.

5. The business English scenario knowledge graph generation and intelligent retrieval method as described in claim 4, characterized in that: The cross-document coreference is softly assigned on the consistency graph. The prior consists of stable reference consistency, triggering semantic similarity and temporal proximity. The graph smoothing term constrains the nearest neighbor event sharing assignment mode. Event clusters are obtained through temperature annealing and threshold hardening. A cluster center vector and consistency score are calculated for each cluster. Isolated event units that fail hardening are returned to the event unit generation stage for re-evaluation and updating of the corresponding evidence anchor index.

6. The business English scenario knowledge graph generation and intelligent retrieval method as described in claim 5, characterized in that: The event chain construction introduces a global order variable and sorts the sequence by intervals. The set of hard constraints on the process limits the irreversible order and participates in the solution together with the weight matrix to obtain an acyclic order. Within the ordered space, directed pairs are selected based on causal posterior. The features of the causal posterior include rank difference, role leader matching, and evidence consistency. Nodes are mapped to discrete state labels using a state decoder. Adjacent high-confidence pairs are merged to form chain segments and summarized into an event chain set.

7. The business English scenario knowledge graph generation and intelligent retrieval method as described in claim 1, characterized in that: The process graph is constructed as a directed acyclic graph based on the candidate edge set, and each node is assigned a hierarchical variable and integerized at output. The inverse hierarchical loss is minimized by combining the prior hierarchy from the process stage and the margin is preserved. Edge-level evidence is aggregated by matching the edge semantic key vector with the anchor text vector according to the evidence anchor index. The edge attributes store the source document identifier, unified timestamp, paragraph offset, and the weight trajectory of the first few high-weight anchors.

8. The business English scenario knowledge graph generation and intelligent retrieval method as described in claim 7, characterized in that: The retrieval intent compilation uses semantic parsing and graph prior jointly to decode and generate query intent expressions. The parsing adopts bundle search and includes violations such as time window, role-object consistency and reachability. The starting point, ending point, necessary and prohibited links are translated into path constraints. The path score integrates causal confidence, constraint satisfaction and time gap penalty, retains multiple candidates and outputs their respective constraint satisfaction and time smoothness indices. The image and text joint execution plan selects evidence for candidate paths under the path constraints. The evidence presentation sub-plan characterizes diversity with a convex combination of semantic similarity kernel and temporal nearest neighbor kernel and adds importance weights, which are limited by evidence budget and terminal type. Anchor points are extracted from the evidence anchor point index and the presentation order is determined to generate a plan object containing paths, evidence lists and presentation instructions.

9. The method for generating and intelligently retrieving business English scenario knowledge graphs as described in claim 8, characterized in that: Intelligent retrieval execution selects the optimal plan by jointly scoring the plan quality function and entropy-temperature risk aggregation. The plan quality function includes three components: path coherence, time smoothing, and evidence credibility. The risk temperature is configurable. The answer posterior is calculated for candidate endpoints in the neighborhood of the optimal path. The answer set for single or multiple endpoints is selected by combining the threshold and the distinctness kernel control. The answer set is then bound to the path and evidence and written into storage.

10. The business English scenario knowledge graph generation and intelligent retrieval method as described in claim 9, characterized in that: The auditable return includes a chained audit trail and de-identified publication. It performs deterministic serialization of the query intent expression, path constraints, optimal plan, answer set, document time sequence index, referential alignment mapping table and evidence anchor index in chronological order and generates collision-resistant hash fingerprints. The serialization includes field sorting, time unit and numerical precision uniformity. The evidence vector applies sparse constraints to the sensitive dimension and retains non-sensitive metadata, and returns a fixed format template.

Images