A Multilingual Large-Scale Dialogue Optimization Method and System Integrating Knowledge Graph

Abstract

This invention discloses a multilingual large-scale model dialogue optimization method and system integrating knowledge graphs, belonging to the fields of natural language processing and artificial intelligence dialogue technology. It addresses the problems of named entity recognition, time-varying attribute processing, and dynamic knowledge graph updates in cross-language dialogues. A multilingual named mention extractor scans each round of dialogue text word-by-word, forming a traceable mention index. Cross-language retrieval and candidate entity localization in the knowledge graph are performed using a multilingual pre-trained semantic model and alias matching paths. For time-varying attributes and title changes, priority graph adjustment is performed based on time slice windows and temporal evidence to ensure a smooth transition between old and new titles. The accuracy and stability of the final dialogue response are ensured by constructing a sparse cost matrix and an uncertainty re-discrimination process. This invention effectively solves the title switching problem in multi-context dialogue systems and improves the application effect of knowledge graphs in dialogue.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and artificial intelligence dialogue technology, and more specifically, to a method and system for optimizing multilingual large-scale model dialogues that integrates knowledge graphs. Background Technology

[0002] With the rapid development of artificial intelligence technology, cross-language dialogue systems are gradually becoming an indispensable part of daily life and business scenarios. Traditional dialogue systems often face challenges in cross-context and cross-cultural situations, especially in handling multilingual named entities, title changes, and time-varying attributes, where they often lack efficient processing mechanisms. While existing methods can perform cross-language retrieval and entity recognition to a certain extent, they often suffer from path dependence, inaccurate candidate entities, and unsmooth temporal transitions. Furthermore, how to efficiently handle knowledge injection and dialogue generation in multilingual scenarios and large-scale knowledge graphs remains a major technological challenge.

[0003] To address the above problems, this invention proposes a solution. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method and system for optimizing multilingual large-scale model dialogues that integrates knowledge graphs, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] In a preferred embodiment, it includes:

[0007] The multilingual naming reference extractor filters out references to people, organizations, and places, binds them with session identifiers, timestamps, and language tags, and then initiates a semantic model and alias matching path for cross-language retrieval.

[0008] Divide time slices into windows based on time expression and semantic change points, read the effective time interval of attributes and combine them with event time anchors to construct a temporal evidence set and determine the comprehensive evaluation value Θ. Then, adjust the candidate costs based on the comprehensive evaluation value Θ, construct a priority graph, explicitly encode the title transition, and apply order-preserving constraints.

[0009] A sparse cost matrix is ​​formed by constructing locally highly correlated candidate blocks. The Sinkhorn iteration is used to solve for the mentioned anchor entities and confidence levels. Unstable candidates are labeled and weighted by an uncertainty re-discrimination process.

[0010] In a preferred embodiment, after receiving cross-language dialogue text, the text is scanned by a multilingual naming reference extractor to filter out references to people, organizations, and places, and these references are bound to conversation identifiers, timestamps, and language tags to form traceable index entries.

[0011] Then, two recall paths are initiated:

[0012] One approach is to calculate the similarity between mentioned text and entity vectors using a semantic model to perform cross-language retrieval;

[0013] Secondly, it can quickly locate information by matching multilingual aliases in the knowledge graph.

[0014] In a preferred embodiment, when the reference and candidate name cross character systems, transliteration or transcription is performed, and the edit distance and n-gram overlap are calculated to obtain a prior score. Then, a graph neighborhood consistency score is calculated for each candidate entity, a time-smoothing prior is injected, and the source and freshness of the candidate entity name are verified. The results are biased and adjusted according to authority and update time, and the evidence is fused to generate the final cost matrix.

[0015] In a preferred embodiment, each round of dialogue is divided into different time slices based on temporal expression and semantic change points. The length and step size of the time slice are selected according to the conversation pattern, and slicing is forced when a change in address or semantic change is detected.

[0016] Then, the effective time interval of the title-related attributes is read from the knowledge graph, and event time anchors are introduced.

[0017] In a preferred embodiment, a temporal evidence set is constructed based on time window, attribute interval, time anchor and version trajectory, and the comprehensive evaluation value Θ of temporal alignment is calculated using this as input. The comprehensive evaluation value Θ is a weighted synthesis of four factors: overlap between attribute interval and window, fit between time anchor and window, monotony of title transition and reproducibility of new version evidence. A high comprehensive evaluation value Θ indicates that the candidate is consistent with the current time context, while a low comprehensive evaluation value Θ indicates that the current time context is inconsistent.

[0018] Then, based on the evaluation results of the comprehensive evaluation value Θ, a priority graph is constructed in each time slice window to sort the candidate entities, adjust the base cost of the candidate entities, and set an evaluation value threshold. If the comprehensive evaluation value Θ is higher than the evaluation value threshold and is consistent with the version trajectory, the base cost is reduced; if the comprehensive evaluation value Θ is lower than the evaluation value threshold, the base cost is increased, thereby correcting the candidate entities of the new version.

[0019] Next, the transition from the old title to the new title is explicitly encoded through the directed relationship across segments, and order preservation constraints are applied. When a user briefly backtracks to the old title, the reverse migration will incur a penalty.

[0020] In a preferred embodiment, the comprehensive evaluation value Θ is used as a driving signal to dynamically adjust the weight of the time smoothing prior; when the comprehensive evaluation value Θ significantly supports the new version, the time smoothing bonus of the old version is reduced; if the comprehensive evaluation value Θ does not significantly support the new version, only the lightweight continuation is retained.

[0021] For cases where the evidence is close, counterfactual verification is adopted. Without changing other evidence, two alignment settings are constructed: old priority or new priority. The evidence gain driven by time anchor and the context consistency gain are compared. If the counterfactual advantage favors the new version, the base cost and time smoothing scaling factor are adjusted simultaneously.

[0022] In the entropy regular optimal transport solution, the weight adjustment driven by the comprehensive evaluation quantity Θ and the counterfactual verification are injected as soft constraints through gating solution. Before each round of alternating scaling of rows and columns, the candidate weights are adjusted according to the comprehensive evaluation quantity Θ and the counterfactual advantage quantity, so that the new version candidate is more likely to be stably allocated, and the old version that is outdated or unsuitable is restricted to the low weight region.

[0023] In a preferred embodiment, as the new version is repeatedly supported in adjacent windows, the confidence accumulation gradually increases the adjustment range of the cost term, and converges to zero if the evidence is insufficient, thus maintaining conservatism.

[0024] In a preferred embodiment, each mention and its candidate set are combined to form a locally highly relevant candidate block, thus forming a sparse cost matrix;

[0025] By using a multilingual pre-trained model, alias hits and high freshness hits combined with out-of-block location masking, and using Sinkhorn iteration to solve the optimal transmission of the sparse matrix, the anchor entity and confidence of each mention are obtained.

[0026] Then, through the uncertainty re-judgment process, candidates with high entropy or insufficient advantages are marked, and those that cannot be clarified are downgraded in weight.

[0027] In a preferred embodiment, the module includes: a mention extraction and cross-language retrieval module, a temporal alignment and priority graph adjustment module, a sparse cost matrix and uncertainty processing module, and a data storage module, with signal connections between the modules.

[0028] The data storage module is used to process all data during the process;

[0029] The mention extraction and cross-language retrieval module is used for temporal alignment and priority graph adjustment. The module is used to filter mentions of people, organizations and places through a multilingual naming mention extractor, bind session identifiers, timestamps and language tags, and start the semantic model and alias matching path for cross-language retrieval.

[0030] The temporal alignment and priority graph adjustment module is used to divide time slice windows according to time expression and semantic change points, read the effective time interval of attributes and combine them with event time anchors to construct a temporal evidence set and determine the comprehensive evaluation quantity Θ. Then, the candidate cost is adjusted according to the comprehensive evaluation quantity Θ, a priority graph is constructed, the title transition is explicitly encoded, and order preservation constraints are applied.

[0031] The sparse cost matrix and uncertainty handling module is used to form a sparse cost matrix by constructing locally highly correlated candidate blocks, use Sinkhorn iterative solution to solve for mentioned anchor entities and confidence levels, and use an uncertainty re-discrimination process to label unstable candidates and perform weight reduction processing.

[0032] The technical effects and advantages of this invention, which integrates a multilingual large-scale model dialogue optimization method and system based on knowledge graphs, are as follows:

[0033] This invention significantly improves the accuracy and efficiency of cross-language retrieval by combining a multilingual naming reference extractor with a semantic model. The joint characterization using time-slice windows and temporal evidence enables the system to accurately handle time-varying attributes such as title changes and job title shifts, ensuring a natural transition for users to the new title during conversations. Based on priority graph adjustment using the comprehensive evaluation metric Θ, the system can dynamically adjust the cost of candidate entities, effectively avoiding the path dependency problem of traditional time-smoothing priors. Simultaneously, the use of block sparse cost matrices and the Sinkhorn iteration method reduces computational load, making large-scale knowledge graph applications more scalable in multilingual scenarios. Through an uncertainty re-judgment process, the system can accurately identify unstable links, avoid introducing erroneous information, and ensure the reliability and consistency of dialogue content. Attached Figure Description

[0034] Figure 1 This invention provides a multilingual large-scale model dialogue optimization method and system timing diagram that integrates knowledge graphs.

[0035] Figure 2 This is a schematic diagram of the multilingual large-scale model dialogue optimization method and system modules that integrates knowledge graphs according to the present invention. Detailed Implementation

[0036] 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.

[0037] Example

[0038] This invention discloses a method for optimizing dialogue in a multilingual large-scale model that integrates knowledge graphs, such as... Figure 1As shown, it includes:

[0039] First, after receiving cross-language dialogue text, the multilingual named mention extractor performs word-by-word scanning and language-adaptive segmentation on each round of text to create an initial screening sequence of mentions of people, organizations, place names, and proper nouns. Then, each mention is immediately bound to its conversation identifier, round timestamp, and language tag in a ternary manner to obtain traceable mentions and context index entries, thereby enabling direct connection of subsequent prior injection, time smoothing, and source annotation.

[0040] The text and language tags mentioned are from user input, the timestamps and session identifiers are from session metadata, and language identification is done through a multilingual text recognizer. All sources are recorded for auditing purposes.

[0041] Next, once the index is ready, two candidate retrieval paths are started in parallel, as follows:

[0042] One approach uses a multilingual pre-trained semantic model to vectorize the multilingual names or descriptions of entities in the mentioned text and knowledge graph, respectively, to create mention vectors and entity vectors in a shared representation space, thereby obtaining cross-language similarity scores to achieve cross-language retrieval of semantically equivalent appellations; the other approach directly reads the multilingual alias field of the knowledge graph records and performs word-based matching based on language tags to obtain a candidate list of alias matches, thereby achieving rapid location of registered aliases.

[0043] The semantic vectors mentioned above are derived from the multilingual model parameters and input text, while the aliases and descriptions are derived from the multilingual fields of the knowledge graph. The source and version of each path hit are registered simultaneously.

[0044] If the mention and candidate name are from different character systems, then the transliteration or transcription is first performed based on the public phonemes, character mappings and language specifications, mapping either one to the other's character domain. Then the edit distance and n-gram overlap are calculated to obtain the string or transliteration prior score. The mapping table and specification used are used as the source of evidence for this prior and are attached to the mention and candidate relationship.

[0045] To ensure verifiability, the mapping tables and standards used for transliteration or transcription are fixed in priority as follows: mainstream target language standards > general academic standards > project-internal supplementary tables; when there is a conflict between two-way mappings, the side with the smaller edit distance is taken and a conflict flag is registered. The prior score for strings or transliterations is recorded as follows: Defined as:

[0046] ；

[0047] To complete the edit distance after transliteration or transcription, κ is the temperature parameter. This prior score, along with the mapping table and specification used, is linked to the corresponding mention-candidate relationship.

[0048] Next, for each candidate entity, the relational adjacency set in the knowledge graph is collected, and the list of anchored or significantly co-occurring entities is aligned within the context of the same session. The neighborhood intersection strength is calculated and recorded as the graph neighborhood consistency score. The anchored entities are from the previous round of alignment output, and the adjacency set is from the knowledge graph triples. If it is the first mention in the current session, the anchored entity is empty.

[0049] Then, the previous round's allocation probability and the current timestamp are read. The matching cost of the high-confidence anchored entity in the previous round is reduced in the current round to inject time-smoothing priors, so as to obtain the cross-round continuation of referential preferences, thereby realizing the continuous continuation of multi-round context; the previous round's distribution and timestamp are respectively derived from the previous round's optimal transmission output and session metadata.

[0050] Subsequently, source and freshness verification was initiated, checking the source reliability and update time of each candidate name and alias, marking the authority and timeliness level, and specifying the bias writing method: based on the base cost. middle Weighting, and prioritizing within the film. Higher-ranking candidates will be penalized with a lower rank.

[0051] Prioritize matches from authoritative sources and those that have been recently updated, and downgrade matches from weak sources or those that have not been updated for a long time; the source identifier comes from the source field and update timestamp of the knowledge entry, and external text time metadata can be added if necessary to strengthen the evidence of freshness.

[0052] Source and Freshness Measurement: Source authority level L∈[0,1]; Freshness score In summary: And write it into the basic cost fusion term.

[0053] Once the evidence is complete and standardized, linear fusion is immediately performed on the five pieces of evidence according to the pre-calibrated weights to calculate the comprehensive cost of each mention and candidate pair and fill it into the corresponding position in the cost matrix. The weighting coefficients are calibrated empirically by leaving room for sessions, for example, with the goal of minimizing the wrong link index, only the batch and effective time are recorded without exposing the specific values. Candidate positions that are not recalled are set as infeasible masking to constrain the search domain.

[0054] It should be noted that in contexts where time-varying attributes are highly sensitive, such as job changes, temporary titles, competition rankings, or provisional appellations, users often smoothly transition from the old appellation to the new one within the same conversational flow. Relying solely on the time smoothing prior and the source or freshness prior in the original solution still has the following two structural shortcomings;

[0055] First, time-smoothing priors are path-dependent when extending across wheels, making it easy to maintain the relative advantage of the anchored entity of the previous wheel due to inertia, and making it difficult to relinquish priority in time at the critical point of switching from old to new.

[0056] Secondly, the prior knowledge of source or freshness measures the strength of credibility by being more recent, but it does not explicitly characterize the alignment between the effective range of candidate entity attributes and the current dialogue time context.

[0057] Ultimately, when the user's true intent has switched to the new version's name, the cost matrix may still maintain a low cost for the old version's candidate. This causes the old and new candidates to form a high-entropy near-parallel structure after the entropy regularization optimal transmission solution, leading to repeated clarifications or even introducing outdated facts into the response chain.

[0058] Therefore, in this embodiment, before entering cost fusion, the current session is first divided into time slices based on the time expression and semantic change points in the dialogue text, so that each mention and its candidates are clearly attached to the corresponding slice window; at the same time, the effective and ineffective time intervals of the attributes related to the title are read from the knowledge graph, and verifiable event points are introduced as time anchors, including but not limited to the official announcement release date, the effective date of registration change, the time of publication of the authoritative list and key nodes of the competition schedule; and the evolution trajectory of the title terms is extracted between adjacent slice windows to depict the directional transition from old to new.

[0059] Window parameters: Window length and step size are selected according to the session type: w∈[1,5] in rounds. Alternatively, the time unit can be w∈[30 seconds, 10 minutes], s∈[0.5w, w]. When a sudden change in terminology or a significant semantic change is detected, a slice is forced at that point; the event time anchor is aligned with the slice window center according to the ±τ tolerance.

[0060] Next, based on the joint characterization of window segments, attribute intervals, time anchors, and version trajectories, a temporal evidence set is formed. This set is then used as input to construct a comprehensive evaluation metric Θ for temporal alignment. Θ is a weighted composite of four normalized components: the degree of overlap between attribute intervals and windows, the strength of fit and causal sequence between time anchors and windows, the monotonous transition of terminology within consecutive windows, and the independent recurrence of new version evidence from heterogeneous sources. A higher Θ value indicates stronger temporal consistency of the candidate within the current window; a lower Θ value indicates insufficient alignment with the current temporal context.

[0061] For ease of implementation, the comprehensive evaluation value Θ of temporal alignment is calculated for mention i and candidate j within the window t as follows:

[0062] ;

[0063] ;

[0064] The value is estimated by the monotonic transition degree of the terms used in adjacent windows, and takes the value [0,1].

[0065] Independent reproducibility of new version evidence from heterogeneous sources.

[0066] τ is the time decay constant.

[0067] After obtaining Θ, an intra-slice priority graph is constructed before the global solution. Candidates within each slice are ranked according to attribute intervals, temporal anchor alignment strength, and consistency with the version trajectory direction. The ranking results are written into the mentions and the candidate's base cost in the form of temporal consistency cost terms. An evaluation threshold is set. When Θ is lower than the evaluation threshold, the overall cost of the candidate within the slice is increased. When Θ is higher than the evaluation threshold and its direction is consistent with the version trajectory, its overall cost is reduced, so that the new version candidate receives structural correction within the slice.

[0068] The temporal consistency cost term corresponding to the intra-slice priority graph is defined within slice window t as:

[0069] The direction and version trajectory are consistent:

[0070] The direction and version trajectory are contrary:

[0071] The method for obtaining the evaluation threshold is as follows:

[0072] The monotonic probability calibration method is used to map Θ to a consistent probability. Then, within the calibration probability domain, with the goal of balancing recall and accuracy, the position that maximizes the difference between the true positive rate and the false positive rate is selected as the optimal cutoff point along the receiver operating characteristic curve. Finally, the optimal cutoff point is used to calculate the evaluation threshold through the calibration function.

[0073] Subsequently, to suppress path dependence of time-smoothing priors, order-preserving constraints are applied between time slices, explicitly encoding the evolution of old to new titles as a directed relationship across slices; and the penalty is reduced for cross-slice migrations in the same direction, while the penalty is increased for reverse migrations; when a user briefly backtracks to refer to an old title, the entry of that candidate is not blocked, but its reverse cross-slice edge is automatically penalized, thereby limiting the continued advantage of the old version in cost fusion and avoiding the long-term suppression of the new version due to the inertia of time smoothing.

[0074] Furthermore, in order to quickly complete the transfer of priority between the old and new versions within the critical window, a priority transfer function is applied to the time smoothing prior using Θ as the driving signal, and Θ is mapped to a dynamic scaling factor of the time smoothing weight: when Θ significantly supports the new version, the time smoothing score of the old version in this round is automatically reduced; when Θ is not significant, only a lightweight continuation is retained to maintain stability.

[0075] The priority transfer function is as follows: ,in Here, represents the original time-smoothing weights, and η is the scaling factor. When When the time smoothing prior of the old version is increased, it is automatically reduced.

[0076] For cases where old or new evidence is presented side-by-side and the evidence is similar, a lightweight counterfactual verification is introduced. Without changing other evidence, two alignment settings are constructed: old priority or new priority. The evidence interpretability gain driven by time anchor and the context consistency gain are compared under the two settings, and the difference is used to define the counterfactual advantage. When the advantage clearly favors the new version, the base cost and time smoothing scaling are immediately fine-tuned synchronously to ensure that the priority is transferred in a timely manner within the window, and the counterfactual advantage is archived along with the trigger window and candidate identifier.

[0077] The counterfactual advantage is defined as follows: ,in This represents the weighted gain of time anchor fit and context consistency under corresponding positional settings. If If so, the overall cost of the new version candidate in the window will be reduced synchronously, and the time smoothing weight of the old version will be reduced according to the priority transfer function.

[0078] When entering the entropy-regular optimal transport solution, a gating method is adopted to inject the above-mentioned pre-process mechanism into the scaling process in the form of soft constraints: before each round of row and column alternating scaling, the candidate weights of the corresponding row are gating adjusted according to the current window's Θ and counterfactual advantage, which is equivalent to applying a mild prior bias to the feasibility probability quality of the row, making it easier for new version candidates that are clearly aligned with the current time context to obtain stable soft assignments, while old version candidates that are outdated or not aligned enough are constrained to the low-weight region; when Θ and counterfactual advantage remain consistent in multiple rounds of scaling, the gating strength naturally decays with iteration, avoiding excessive squeezing of the final solution.

[0079] The gating weight is denoted as .

[0080] Overall Cost Corresponding OT core .

[0081] Retrieve after gating:

[0082] :

[0083] ;

[0084] ;

[0085] The gating strength decreases geometrically with a fixed step size during the iteration process.

[0086] If, within all the window segments, any candidate attribute range lacks effective alignment with the temporal context and has insufficient temporal anchor support, then the relative absorption capacity of NIL is improved, allowing the mention to naturally flow towards the unknown entity during the solution process, avoiding being forced to fit to outdated or incompatible versions.

[0087] Subsequently, to adapt to cross-topic and long conversations, a reliable cumulative amount is maintained at the slice window level that rolls over time to balance short-term noise and medium-term trends: when the new version repeatedly obtains joint support of attribute ranges and time anchors in adjacent slice windows, the reliable cumulative amount gradually increases the adjustment range of cost terms and time smoothing scaling; when only occasional text mentions occur and are not confirmed by the dual verification of in-slice priority graph and counterfactual verification, the adjustment range automatically converges to zero to maintain conservatism.

[0088] Meanwhile, for each mention, only its locally highly relevant candidate blocks are retained to form a block sparse cost matrix, so as to obtain a near-linear effective computational domain, providing a scalable numerical foundation for subsequent global alignment in large-scale knowledge graphs and multilingual scenarios.

[0089] Block Construction and Complexity: Top-M candidates (M ranges from 10 to 100) from the multilingual pre-trained semantic model, along with word-matter hits and high-freshness hits, are merged into local candidate blocks, with out-of-block positions masked. Let Z be the number of non-zero entries, then the complexity of each step of Sinkhorn is approximately O(Z). With a fixed block sparsity, the overall computational load increases approximately linearly with Z.

[0090] It should be noted that the value of any generation on this matrix can be traced back to the corresponding cross-linguistic semantics, string or transliteration, graph neighborhood, time smoothing and original evidence of source or freshness, calculation method and source time, to avoid situations where no source data is involved in the decision-making.

[0091] After the cost matrix is ​​formed, an additional column is added to represent NILs without corresponding entities. The placeholder column is regarded as an optional allocation destination with equal status to other candidates, in order to explicitly absorb emerging names or unregistered names outside the knowledge base and avoid forced incorrect fitting during global alignment.

[0092] NIL cost settings:

[0093] NIL determination: When the probability of NIL in this row is... When this happens, the mention will be marked as an unknown entity and its proprietary knowledge injection will be suspended.

[0094] After the cost domain and candidate domain are prepared, the block sparse cost matrix is ​​solved by applying entropy regularization to solve the optimal transmission using the entire session as the solution unit. The Sinkhorn iteration is used to perform alternating scaling of rows and columns until the row constraints are satisfied and the column-side soft capacity reaches convergence. Infeasible locations are always kept shielded during the iteration process to prevent illegal traffic leakage.

[0095] Subsequently, the output mentions and candidate probability coupling matrix are solved. For each mention, the maximum probability term in its row is read as the anchor entity, and the maximum probability is recorded as the anchor confidence. When the probability term corresponding to the NIL column is the maximum and exceeds an empirical threshold, such as above an empirical threshold line, the mention is marked as an unknown entity and its proprietary knowledge injection is suspended, so as to cut off potential error tracing at the output level.

[0096] Uncertainty re-discrimination threshold: line entropy Advantage difference .when or When this occurs, the mention will be added to the uncertain set to trigger clarification or downweighting.

[0097] To maintain strict parameter interpretability, only the necessary objective expression is given here without lengthy derivations: Optimal transport is minimized with respect to the expected total cost and the entropy regularization term, denoted as... Let i be the probability of the i-th mention being assigned to the j-th candidate containing NIL. Given the corresponding overall cost, γ is the entropy regularization coefficient, and the objective can be written as minimizing... in This is used to ensure the conservation of single mention probability and to express column-side capacity in the form of soft constraints.

[0098] in, Derived from the output of the solution iteration, The cost fusion is derived from the aforementioned cost fusion, where γ is derived from the verification session calibration and registration batch and effective time. Then, through the combination of block sparse candidates and Sinkhorn iteration, the computational load increases approximately linearly with the number of non-zero entries.

[0099] Optimal transmission target: Minimize

[0100] Constraint: For each i, there is The column side is represented by soft capacity b. Take... In gating Under the influence of, Next, the Sinkhorn iteration is used: , Infeasible locations are kept masked; after the solution is completed, the anchored entity and confidence level are determined by the largest term in each row.

[0101] Furthermore, after obtaining the anchoring results and confidence distribution, the uncertainty re-judgment process is immediately initiated to identify and address potential unstable links, as follows:

[0102] First, calculate the row entropy of each mentioned row based on the coupling matrix, and mark the high-entropy entries with dispersed probabilities; then calculate the interval between the probabilities of the first and second place, and mark the nearly tied entries with insufficient advantages.

[0103] Subsequently, a baseline method decoupled from the main process is run independently, such as generating Top-1 results using only string rules or another set of cross-language embedding encoders, calculating the consistency with the main results and marking dissenting opinions; if any mark is hit, the mention is added to the uncertain set and a disposition instruction is generated.

[0104] If the dialogue context allows for interaction, a clarifying multiple-choice statement is constructed on the spot, listing several high-scoring candidate canonical names and the shortest identification information for the user to choose from, and the selection result is overwritten with the alignment record of the mention; if the context is not suitable for clarification or needs to be output immediately, the proprietary knowledge of the entity involved in the mention is downgraded, and the entity-specific facts are temporarily replaced with conditional or general descriptions, and the mention is registered as pending secondary confirmation.

[0105] Among them, the line entropy index comes from the OT output distribution, the gap between the top two comes from the ranking difference of the same distribution, the integration consistency comes from the baseline output that is completely independent of the main process, and the clarification write-back comes from the user's selected text.

[0106] While uncertainty handling is carried out simultaneously, an alias generation and verification loop is immediately triggered for all high-confidence anchor entries to expand alias knowledge in a safe and verifiable manner.

[0107] Specifically, the name actually used by the user in this round is first written directly into the candidate alias pool to form the first batch of candidates; then, through controlled multilingual generation, only name-level variants are generated, such as uppercase and lowercase variants, common abbreviations or literal translations. The generated results are added to the candidate alias pool one by one and the model, hints and version information used are recorded.

[0108] Next, each candidate alias was verified and scored based on four types of independent evidence:

[0109] First, search the verifiable text for the co-occurrence of candidate aliases and target entities in terms of domain terms, related people and key attributes, record the source and form graph neighborhood consistency evidence;

[0110] Secondly, the co-occurrence frequency of the alias and the target entity is counted in historical conversations and corpora on the same topic. If the user has explicitly confirmed the clarification in this round, it is directly counted as strong evidence, thus forming dialogue co-occurrence evidence.

[0111] Third, calculate the edit distance and n-gram overlap between the alias and the known name or alias of the entity. If necessary, perform transliteration mapping first and then compare to form evidence of character or pronunciation similarity, and record the mapping table and the source of language norms.

[0112] Fourth, machine translation is invoked to translate the alias back into the entity's main language and compare it with the main name or main alias (or reverse translation to verify consistency), and the translation tool and version are recorded to form cross-language back-translation consistency evidence.

[0113] Afterwards, the four pieces of evidence are normalized and combined into a comprehensive score. When the comprehensive score exceeds the activation threshold, such as a low threshold line, the alias is activated in the current session and written to the shadow library. When the alias appears multiple times in subsequent independent sessions and the cumulative comprehensive score exceeds the confirmation threshold, such as a high threshold line, it is written back to the knowledge graph main alias table. If the confirmation threshold is not reached for a long time, it is retained in the shadow library and decayed or removed over time.

[0114] Furthermore, during the continuous operation of the closed loop, the response generates factual content that is only injected with high-confidence anchored entities, and the features of entities that have been marked as uncertain are uniformly masked; once the clarification write-back arrives, the anchor of the mention is refreshed immediately and the subsequent time-smoothed prior is updated; when the user switches languages ​​or titles in the same session or subsequent sessions, the aliases already enabled in the shadow library, as well as the existing transliteration and back-translation record paths and source annotations, are used to ensure semantic consistency and alignment stability of cross-language expressions.

[0115] Subsequently, at the end of the session, a session-level aligned summary is generated, which summarizes the anchored entities, NIL judgment entries, uncertain handling records and alias closure status, and archives the weight batches, threshold examples, evidence numbers and tool versions in a unified manner to form a complete audit trail.

[0116] The cross-session phase maintains a block sparse candidate strategy to limit the computation domain, prioritizing cross-language semantic hits and high-freshness source hits while eliminating duplicate or low-reliability candidates; in the optimal transmission phase, the scaling factor of the previous session is reused as initialization to shorten the Sinkhorn convergence steps; in alias management, the shadow library is scanned periodically to merge synonym variants and clean up long-term low-reliability aliases; in weight and threshold management, the weighting coefficients and trigger thresholds are empirically calibrated based on the verification session, and the effective batches and time points are recorded to avoid introducing unsourced parameter drift, thereby maintaining near-linear expansion, cross-language consistency, and long-term stability under the conditions of scale growth and domain migration.

[0117] This invention also proposes a multilingual large-scale model dialogue optimization system that integrates knowledge graphs, such as... Figure 2 As shown, it includes: a mention extraction and cross-language retrieval module, a temporal alignment and priority graph adjustment module, a sparse cost matrix and uncertainty handling module, and a data storage module, with signal connections between the modules;

[0118] The data storage module is used to process all data during the process;

[0119] The mention extraction and cross-language retrieval module is used for temporal alignment and priority graph adjustment. The module is used to filter mentions of people, organizations and places through a multilingual naming mention extractor, bind session identifiers, timestamps and language tags, and start the semantic model and alias matching path for cross-language retrieval.

[0120] The temporal alignment and priority graph adjustment module is used to divide time slice windows according to time expression and semantic change points, read the effective time interval of attributes and combine them with event time anchors to construct a temporal evidence set and determine the comprehensive evaluation quantity Θ. Then, the candidate cost is adjusted according to the comprehensive evaluation quantity Θ, a priority graph is constructed, the title transition is explicitly encoded, and order preservation constraints are applied.

[0121] The sparse cost matrix and uncertainty handling module is used to form a sparse cost matrix by constructing locally highly correlated candidate blocks, use Sinkhorn iterative solution to solve for mentioned anchor entities and confidence levels, and use an uncertainty re-discrimination process to label unstable candidates and perform weight reduction processing.

[0122] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0123] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0124] Those skilled in the art will recognize that the modules 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 inventive 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.

[0125] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0126] 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.

[0127] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for optimizing multi-language large model dialogue based on a fusion knowledge graph, Characterized in that; comprising: Filtering out the mentions of person, organization and place through the multilingual named entity extractor, binding the session ID, timestamp and language label, starting the semantic model and alias matching path for cross-language retrieval; According to the time expression and semantic change point, the time slice window is divided, the attribute effective time interval is read and combined with the event time anchor to construct the temporal evidence set, and the comprehensive evaluation quantity Θ is determined, then the candidate cost is adjusted according to the comprehensive evaluation quantity Θ, the priority graph is constructed, the transition of address is explicitly coded, and the order preserving constraint is applied; Each round of dialogue is divided into different time slice windows according to the time expression and semantic change point, wherein the length and step of the time slice window are selected according to the session form, and when the address mutation or semantic change is detected, the slicing is forced to be performed; Then, the attribute effective time interval related to the address is read from the knowledge graph, and the event time anchor is introduced; Based on the time slice window, the attribute interval, the time anchor and the version track, the temporal evidence set is constructed, and the comprehensive evaluation quantity Θ of the temporal alignment is calculated, wherein the comprehensive evaluation quantity Θ is composed of the attribute interval and the slice window overlap, the time anchor and the slice window fitting degree, the monotonicity of the address transition and the recurrence degree of the new version evidence, and the comprehensive evaluation quantity Θ value is high, indicating that the candidate is consistent with the current time context, and the comprehensive evaluation quantity Θ value is low, indicating that the current time context is inconsistent; Then, based on the evaluation result of the comprehensive evaluation quantity Θ, the priority graph is constructed in each time slice window, the candidate entity is sorted, the basic cost of the candidate entity is adjusted, and the evaluation quantity threshold is set, if the comprehensive evaluation quantity Θ value is higher than the evaluation quantity threshold and consistent with the version track, the basic cost is reduced; if the comprehensive evaluation quantity Θ value is lower than the evaluation quantity threshold, the basic cost is increased, thereby supporting the new version candidate entity; Then, the transition from the old address to the new address is explicitly coded through the cross-slice directed relationship, and the order preserving constraint is applied, when the user short-time backtracks to the old address, the reverse migration will increase the penalty; By constructing a local high correlation candidate block to form a sparse cost matrix, Sinkhorn iteration is used to solve the reference anchor entity and confidence, and through the uncertainty re-identification process, unstable candidates are marked and weighted down.

2. The method of claim 1, wherein the method is performed by a server. After receiving the cross-language dialogue text, the text is scanned by the multilingual named entity extractor, the mentions of person, organization and place are filtered out, and the session ID, timestamp and language label are bound to form traceable index entries; Then, two recall paths are started: One is to calculate the similarity between the mention text and the entity vector through the semantic model for cross-language retrieval; The second is to quickly locate by matching the multilingual aliases in the knowledge graph.

3. The method of claim 2, wherein the method is characterized by: When the mention and the candidate name cross the character system, the transliteration or transcription is performed, the edit distance and n-gram overlap are calculated to get the prior score, then the graph neighborhood consistency score is calculated for each candidate entity, the time smoothing prior is injected, and the source and freshness of the candidate entity name are verified, the results are biased and adjusted according to the authority and update time, and the final cost matrix is fused with the evidence.

4. The method of claim 3, wherein the method is characterized by: The comprehensive evaluation quantity Θ is taken as a driving signal to dynamically adjust the weight of the time smoothing prior. When the comprehensive evaluation quantity Θ significantly supports the new version, the time smoothing bonus of the old version is reduced. If the comprehensive evaluation quantity Θ does not significantly support the new version, only the lightweight continuation is retained. For the case of evidence proximity, counterfactual verification is adopted. Without changing other evidence, two kinds of opposite settings, old priority or new priority, are constructed. The time anchor driven evidence gain and the context consistency gain are compared. If the counterfactual advantage quantity tends to the new version, the basic cost and the time smoothing scaling coefficient are simultaneously fine-tuned. In the entropy regularized optimal transport solution, the weight adjustment driven by the comprehensive evaluation quantity Θ and the counterfactual verification are injected as soft constraints through gated solution. Before each round of row and column scaling, the candidate weight is adjusted according to the comprehensive evaluation quantity Θ and the counterfactual advantage quantity, so that the new version candidate can be more easily allocated stably, and the outdated or unsuitable old version is limited to the low weight area.

5. The method of claim 4, wherein the method is characterized in that: When the new version is repeatedly supported within the adjacent slice window, the credible accumulation gradually expands the adjustment range of the cost term. If the evidence is insufficient, it converges to zero, maintaining the conservatism.

6. The method of claim 5, wherein the method is characterized in that: Each mention is formed into a local high correlation candidate block with its candidate set, and a sparse cost matrix is formed. Through the multi-language pre-training model, alias hit and high freshness hit merging block outer position shielding, and using Sinkhorn iteration to solve the optimal transport of the sparse matrix, the anchor entity and the confidence of each mention are obtained. Then, through the uncertainty re-identification process, the candidates with high entropy or insufficient advantage are labeled. If it cannot be clarified, the weight is reduced.

7. A multi-language large model dialogue optimization system fused with a knowledge graph, characterized in that, It includes: The mention extraction and cross-language retrieval module, the time alignment and priority graph adjustment module, the sparse cost matrix and uncertainty processing module, and the data storage module. The signals between the modules are connected. The data storage module is used for all data in the processing process. The mention extraction and cross-language retrieval module is used for the time alignment and priority graph adjustment module to filter out the mentions of characters, institutions and place names through the multi-language named mention extractor, bind the session identifier, timestamp and language label, start the semantic model and alias matching path for cross-language retrieval. The time alignment and priority graph adjustment module is used to divide the time slice window according to the time expression and semantic change point, read the attribute effective time interval, combine the event time anchor, construct the time evidence set, and determine the comprehensive evaluation quantity Θ. Then, the candidate cost is adjusted according to the comprehensive evaluation quantity Θ, the priority graph is constructed, the title transition is explicitly coded, and the order preserving constraint is applied. Each round of dialogue is divided into different time slice windows according to the time expression and semantic change point. The length and step of the time slice window are selected according to the session form, and when the title mutation or semantic change is detected, the slicing is forced. Then, the attribute effective time interval related to the title is read from the knowledge graph, and the event time anchor is introduced. A temporal evidence set is constructed based on a time slice window, an attribute interval, a time anchor and a version track, and a comprehensive evaluation quantity Θ of temporal alignment is calculated based on the temporal evidence set as input, wherein the comprehensive evaluation quantity Θ is weighted and synthesized by four factors including attribute interval and slice window overlap, time anchor and slice window fit degree, monotonicity of appellation transition and recurrence degree of new version evidence, a high value of the comprehensive evaluation quantity Θ indicates that the candidate is consistent with the current time context, and a low value of the comprehensive evaluation quantity Θ indicates that the current time context is inconsistent; Then, based on the evaluation result of the comprehensive evaluation quantity Θ, a priority graph is constructed in each time slice window, the candidate entity is sorted, the basic cost of the candidate entity is adjusted, and an evaluation quantity threshold is set, if the value of the comprehensive evaluation quantity Θ is higher than the evaluation quantity threshold and is consistent with the version track, the basic cost is reduced, and if the value of the comprehensive evaluation quantity Θ is lower than the evaluation quantity threshold, the basic cost is increased, thereby supporting the new version candidate entity; Then, the transition from the old appellation to the new appellation is explicitly coded through the directed relationship across the slices, and the order preserving constraint is applied, when the user backtracks to the old appellation for a short time, the reverse migration will increase the penalty; The sparse cost matrix and uncertainty processing module is used for forming a sparse cost matrix by constructing a local high correlation candidate block, using Sinkhorn iteration to solve the reference anchor entity and the confidence, and through an uncertainty re-identification process, unstable candidates are marked and weight processing is performed.

Images