A method and system for training a korean dialogue state tracking model

Abstract

The present application relates to the technical field of natural language processing and intelligent dialogue, more specifically, the present application relates to a training method and system of a Korean dialogue state tracking model, the method faces a translation auxiliary scene, establishes a target business domain set, a Korean native slot system, a slot dependency relationship graph and a state transition rule graph, constructs a round level sample containing dialogue states before and after updating, generates a slot update difference mask and an update type label, and combines light form analysis, honorific identification, auxiliary word anchor point annotation, omission recovery, slot level evidence chain construction, evidence consistency evaluation and selective write-back mechanism to jointly train the Korean dialogue state tracking model. The scheme can improve the accuracy, stability and robustness of state tracking in Korean multi-round dialogue scenarios.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and intelligent dialogue technology, and more specifically, to a training method and system for a Korean dialogue state tracking model. Background Technology

[0002] With the development of applications such as cross-language intelligent customer service, Korean-language life service Q&A, and Korean-language navigation and booking assistants, state tracking technology for multi-turn Korean dialogues has gradually become a key foundational capability in dialogue systems. The task of dialogue state tracking is to continuously identify the slots and their values ​​related to the user's current intent during multi-turn interactions, and to maintain a dialogue state that can be used for subsequent retrieval, recommendation, confirmation, and generation.

[0003] Most existing dialogue state tracking methods are built around English or other common language scenarios. Their slot systems, semantic boundaries, and update logic are usually based on general ontology, lacking targeted modeling for honorific changes, particle dependencies, ellipsis, and the semantic structure of local services in Korean scenarios. Especially in high-frequency local Korean scenarios such as restaurant reservations, travel navigation, and utility payments, user expressions often simultaneously involve changes in politeness level, slot ellipsis, local corrections, and cross-turn transitions. Relying solely on general state tracking models can easily lead to inaccurate slot identification, delayed state updates, or incorrect update directions.

[0004] Furthermore, existing technologies typically organize training samples by directly mapping the current round's input to the target state, lacking explicit modeling of the evolutionary relationship between the "pre-update state and post-update state" and also lacking focused supervision of rounds where real state changes occur. This results in the model being able to learn the static slot value distribution, but struggling to accurately identify which rounds saw additions, deletions, or replacements, leading to unclear state update boundaries in multi-round dialogues and reduced tracking stability in complex dialogue scenarios.

[0005] Furthermore, Korean dialogues are rife with phenomena such as subject omission, object omission, and short response corrections. Existing technologies employ two main approaches: one directly tracks the state from the original text, which is prone to missing slot values ​​due to missing key information; the other involves rewriting the entire sentence for recovery, but this method easily introduces new semantic shifts and propagates recovery errors into subsequent state updates. Meanwhile, in translation-assisted scenarios, existing technologies typically use the translation results as a uniform supplementary input, lacking uncertainty control for candidate slots and multi-source confidence fusion mechanisms, leading to overly broad or distorted use of auxiliary information.

[0006] Therefore, we propose a training method and system for a Korean dialogue state tracking model to solve the above problems. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a training method and system for a Korean dialogue state tracking model to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a training method for a Korean dialogue state tracking model, comprising the following steps: Step 1: Establish a set of target business domains, a native Korean slot system, a slot dependency graph, and a state transition rule graph for Korean multi-turn dialogue in translation assistance scenarios. The native Korean slot system includes a set of slots corresponding to each target business domain, a set of allowed values ​​corresponding to each slot, and a set of sensitive slots associated with honorific expressions or native Korean service scenarios. The state transition rule graph includes business constraint relationships, temporal constraint relationships, and mutual exclusion constraint relationships. Step 2: Organize the Korean multi-turn dialogue training corpus into a turn-level sample sequence, and construct each turn training sample as a seven-tuple sample including dialogue identifier, turn identifier, speaker identifier, original Korean text of the current turn, dialogue state before update, dialogue state after update, and key turn identifier; Based on the slot-by-slot comparison results between the dialogue state before update and the dialogue state after update, generate a slot update differential mask and a slot update type label. Specifically, when the slot value of a certain slot is different in the dialogue state before update and the dialogue state after update, or one is empty while the other is not empty, the differential mask corresponding to that slot is determined to be the updated state; otherwise, it is determined to be the maintain state. Step 3: Perform Korean preprocessing on the original Korean text for the current round. Korean preprocessing includes at least light morphological analysis, honorific recognition, and particle anchor labeling. The original Korean text character sequence and word order are preserved, and only morphological labels, honorific level labels, and particle anchor position labels are added. Particle separation and replacement and word ending cutting and replacement are not performed. Step 4: Based on the slot update differential mask, the original Korean text of the current round, the dialogue state before the update, and the historical round information in the candidate window, perform the omission recovery trigger judgment; when the omission trigger score after weighting the subject omission indicator, object omission indicator, and partial omission indicator according to their respective weights is not lower than the preset recovery threshold, omission recovery is triggered; generate a candidate recovery set from the noun phrases, confirmed slot values, and dependent slot values ​​adjacent to the suspected update slot in the most recent two to four historical rounds, and select the target candidate to form the recovery text, whereby the recovery text is only used to supplement the semantic missing part corresponding to the suspected update slot; Step 5: Construct a slot-level evidence chain for each suspected update slot. The slot-level evidence chain shall include at least three of the following: original text evidence fragments from the original Korean text, restored evidence fragments from the restored text, particle anchor evidence, honorific level evidence, and translation auxiliary evidence. Among them, the translation auxiliary evidence shall come from at least two of the following: parallel corpus sources, target domain slot mapping dictionary sources, and translation model sources. The translation auxiliary evidence shall be included in the slot-level evidence chain only when the uncertainty of the suspected update slot is higher than the preset uncertainty threshold and the corresponding slot-level fusion confidence is not lower than the preset auxiliary threshold. Step 6: Input the original Korean text of the current round, the restored text, the morphological label, the honorific level label, the particle anchor position label, the dialogue state before the update, the historical round information in the historical context window, the slot update differential mask, and the slot-level evidence chain into the Korean dialogue state tracking model, and output the slot existence result, slot value prediction result, slot update gating result, state transition legality result, restoration candidate selection result, and slot evidence consistency result. Step 7: Based on the slot update gating results, state transition legality results, and slot evidence consistency results, perform selective write-back on the candidate updated dialogue states. Specifically, only slots that simultaneously meet the update gating conditions, state transition rule graph constraints, and evidence consistency threshold conditions are written back. Slots that do not meet any of the conditions retain the original state values ​​in the dialogue states before the update and are recorded as abnormal samples. Step 8: Construct a joint loss function based on slot existence results, slot value prediction results, state transition legality results, recovery candidate selection results, slot update differential mask, and slot evidence consistency results. Then, use the joint loss function to update the parameters of the Korean dialogue state tracking model to obtain the trained Korean dialogue state tracking model. The joint loss function consists of the following weighted sums: slot existence classification loss multiplied by the first weight coefficient, slot value prediction loss multiplied by the second weight coefficient, state transition legality judgment loss multiplied by the third weight coefficient, recovery candidate selection loss multiplied by the fourth weight coefficient, slot update differential consistency loss multiplied by the fifth weight coefficient, and slot evidence consistency loss multiplied by the sixth weight coefficient. Specifically, when the confidence level of slot prediction is lower than the preset threshold, the legality result of state transition is characterized as illegal transition, the highest score of recovery candidate is lower than the preset recovery threshold, the prediction result of slot update differential mask is inconsistent with the labeling result, or the consistency result of slot evidence is lower than the preset evidence threshold, the corresponding training samples are written into the hard sample library and participate in retraining in subsequent training rounds with a higher sampling weight than ordinary samples.

[0009] In a preferred embodiment, the target business domain set includes at least three of the following: restaurant reservation, travel navigation, movie ticket reservation, express delivery tracking, and utility bill payment; wherein, the slot set of the restaurant reservation business domain includes at least three of the following: “음식종류”, “맵기수준”, “음식가격대”, “요청사항”, and “예약자성별”; wherein, the slot corresponding to “예약자 (gender)” is configured as a polite language sensitive slot, used to establish a mapping relationship with changes in Korean titles, changes in politeness levels, or changes in address templates.

[0010] In a preferred embodiment, the slot update type label includes at least three categories: addition, deletion, replacement, and retention. The addition type corresponds to slots that were empty in the pre-update dialogue state and not empty in the post-update dialogue state; the deletion type corresponds to slots that were not empty in the pre-update dialogue state and were empty in the post-update dialogue state; the replacement type corresponds to slots that were not empty in both the pre-update and post-update dialogue states but had different slot values; and the retention type corresponds to slots that had the same slot value in both the pre-update and post-update dialogue states. Furthermore, the key round identifier is derived from the number of slots that were added, deleted, or replaced, and is used to improve the sample supervision strength for actual slot update rounds.

[0011] In a preferred embodiment, the particle anchor position label is obtained by marking the position of at least one particle among “이 / 가”, “을 / 를”, “은 / 는”, “에”, “에서”, and “(으)로” without changing the word segmentation boundary of the original Korean text; the honorific level label is obtained by identifying Korean terminator endings and subject honorific markers; and the particle anchor position label is used to limit the candidate evidence boundary of location slots, time slots, or object slots, while the honorific level label is used to limit the update credibility of request strength-related slots or title-related slots.

[0012] In a preferred embodiment, the original text evidence fragment in the slot-level evidence chain is obtained by performing slot anchor point localization on the original Korean text of the current round, the recovered evidence fragment is obtained by performing same-slot localization on the recovered text, the particle anchor point evidence consists of the particle position and particle type corresponding to the candidate slot, the honorific level evidence consists of the honorific level label of the original Korean text of the current round, and the translation auxiliary evidence consists of the translation auxiliary text corresponding to the candidate slot and its source confidence. Furthermore, the slot evidence consistency result is obtained by performing consistency evaluation on at least three types of evidence among the original text evidence fragment, the recovered evidence fragment, the particle anchor point evidence, the honorific level evidence, and the translation auxiliary evidence. The more valid evidence participating in the consistency evaluation and the more consistent the support direction of the various types of evidence for the update conclusion of the same slot, the higher the evidence consistency result of the corresponding slot.

[0013] In a preferred embodiment, a slot-level fusion confidence score is calculated for each suspected update slot. The slot-level fusion confidence score is obtained by weighting and summing the basic confidence scores provided by each translation assistance source for the suspected update slot according to the fusion weights corresponding to their respective historical accuracy. The sum of each fusion weight is one. Furthermore, the corresponding translation assistance evidence is allowed to participate in the evidence consistency assessment of the suspected update slot only when the uncertainty of the corresponding suspected update slot is higher than a preset uncertainty threshold and the slot-level fusion confidence score is not lower than a preset assistance threshold.

[0014] In a preferred embodiment, the first weight coefficient in the joint loss function ranges from 0.5 to 1.5, the second weight coefficient ranges from 0.8 to 2.0, the third weight coefficient ranges from 0.2 to 0.8, the fourth weight coefficient ranges from 0.3 to 1.0, the fifth weight coefficient ranges from 0.2 to 0.8, and the sixth weight coefficient ranges from 0.2 to 1.0. Furthermore, a hard sample weight of 1.5 to 3.0 is applied to at least one of the following samples: samples with misjudged honorific level, samples with slot extraction errors caused by particle ambiguity, samples with failed omission recovery, samples with illegal state transitions, samples with errors in slot update differential mask prediction, and samples with slot evidence consistency results lower than a preset evidence threshold.

[0015] In a preferred embodiment, a training system for a Korean dialogue state tracking model includes: The slot system construction module is used to establish the target business domain set, the native Korean slot system, the slot dependency graph, and the state transition rule graph; The sample organization module is used to organize Korean multi-turn dialogue training corpus into seven-tuple samples; The differential label generation module is used to generate a slot update differential mask and a slot update type label based on the dialogue state before and after the update. The Korean preprocessing module is used to output morphological tags, honorific level tags, and particle anchor point position tags; The omission recovery module is used to perform omission recovery trigger judgment and generate recovery text; The evidence chain construction module is used to build a slot-level evidence chain for each suspected update slot, including original text evidence fragments, restored evidence fragments, auxiliary word anchor point evidence, honorific level evidence, and translation auxiliary evidence. The evidence consistency assessment module is used to output the evidence consistency results for each slot. The joint training module is used to output slot existence results, slot value prediction results, slot update gating results, state transition legality results, and recovery candidate selection results. The selective write-back module is used to perform selective write-back of candidate states based on the slot update gating result, the state transition legality result, and the slot evidence consistency result. The hard sample closed-loop module is used to write abnormal samples into the hard sample library and participate in subsequent retraining with a higher sampling weight than ordinary samples.

[0016] The technical effects and advantages of this invention are as follows: 1. This invention establishes a target business domain set, a native Korean slot system, a slot dependency graph, and a state transition rule graph around the translation assistance scenario. It combines light morphological analysis, honorific recognition, and particle anchor annotation to perform targeted modeling of politeness level changes, particle dependencies, and service semantics in native Korean scenarios, thereby improving the accuracy of slot recognition and state tracking in Korean multi-turn dialogues.

[0017] 2. This invention constructs round-level samples containing the dialogue state before the update, the dialogue state after the update, and key round identifiers, and generates slot update differential masks and update type labels, enabling the model to explicitly learn the state evolution process, update boundaries, and true update rounds, thereby improving the stability of state change recognition such as addition, deletion, and replacement.

[0018] 3. This invention adopts a trigger-based local supplementation strategy for omission recovery and limits the recovered text to supplement only the semantically missing parts corresponding to the suspected update slots; at the same time, it sets a controlled introduction mechanism for translation assistance evidence based on slot uncertainty and fusion confidence, thereby reducing semantic drift caused by whole sentence rewriting and indiscriminate auxiliary input, and improving the reliability of state tracking in complex omission scenarios and translation assistance scenarios.

[0019] 4. This invention constructs a slot-level evidence chain for each suspected update slot, and performs consistency assessment based on at least three types of evidence, including original text evidence, recovery evidence, auxiliary word anchor evidence, honorific level evidence, and translation auxiliary evidence. Then, it performs selective write-back by combining the update gating result and the state transition legality result, thereby reducing the probability of erroneous write-back and improving the interpretability and controllability of the state update process.

[0020] 5. This invention constructs a joint loss function that includes slot existence classification loss, slot value prediction loss, state transition legality discrimination loss, recovery candidate selection loss, differential consistency loss, and evidence consistency loss. Combined with a hard sample library and a weighted retraining mechanism, it continuously strengthens high-risk samples such as honorific misjudgment, particle ambiguity, omission recovery failure, and illegal transition, thereby improving the robustness and generalization ability of the model in complex Korean multi-turn dialogue scenarios. Attached Figure Description

[0021] Figure 1This diagram illustrates the overall process of training a Korean dialogue state tracking model according to an embodiment of the present invention.

[0022] Figure 2 This illustration shows a flowchart of the process of organizing Korean multi-turn dialogue training corpus into turn-level samples and generating slot update differential masks and update type labels in an embodiment of the present invention.

[0023] Figure 3 This illustration shows a flowchart of the process of performing light morphological analysis, honorific recognition, particle anchoring, omission recovery trigger judgment, and recovery text generation on the original Korean text of the current round in an embodiment of the present invention.

[0024] Figure 4 This illustration shows a flowchart of the process of constructing a slot-level evidence chain, performing evidence consistency assessment, and selectively writing back the state for a suspected update slot in an embodiment of the present invention.

[0025] Figure 5 A schematic diagram of the module structure of a Korean dialogue state tracking model training system according to an embodiment of the present invention is shown. Detailed Implementation

[0026] 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.

[0027] Reference Figure 1-5 A training method for a Korean dialogue state tracking model includes the following steps: Step one involves selecting business domains from high-frequency service scenarios in South Korea that exhibit stable slot structures, a clear need for maintaining state across interactions, and frequent omissions, changes in honorifics, and reliance on particles in user expressions. At least three of the following are prioritized as training target domains: restaurant reservations, travel navigation, movie ticket bookings, courier tracking, and utility bill payments. The reason for using multi-domain joint modeling instead of single-domain modeling is that in translation assistance scenarios, users often switch between domains to express their intent during continuous interactions. Constructing the state space based solely on a single domain can easily lead to incomplete slot transitions and broken context states.

[0028] After identifying the target business domains, a native Korean slot system is established for each domain. Instead of directly using the general English DST ontology, the system is based on the actual expressions used in Korean service scenarios. Slots are defined by semantic variables that users frequently express explicitly or implicitly depend on context in Korean, and a set of allowed values ​​is configured for each slot. This set of allowed values ​​can be pre-defined using a closed enumeration method or constructed using an enumeration value plus open text supplementation, thus balancing high-frequency standard values ​​with free expression in real-world scenarios. Taking the restaurant reservation business domain as an example, at least the following slots can be set: “음식종류”, “맵기수준”, “음식가격대”, “요청사항”, and “예약자성별”. Among them, “맵기수준” is set as a separate slot because the expression of spiciness in Korean restaurant conversations is high-frequency, granular, and directly affects the recommendation results. “예약자성별” is configured separately because Korean titles, honorifics, and descriptions of service recipients are affected by gender-related expressions. This slot is not a general population attribute record, but is used to support the generation of subsequent address templates and politeness levels.

[0029] After the slot system is established, a slot dependency graph is further constructed. The construction principle is: if the legal value range, update priority, or state interpretation of one slot depends on another slot, then a dependency edge is established between them. For example, in the restaurant reservation business domain, there is a dependency relationship between "맵기수준" and "음식종류" because different cuisines may have different allowed ranges for spiciness values; in the travel navigation business domain, there are also paired constraints between the departure point and the arrival point. By pre-establishing the slot dependency graph, the context search range of candidate update slots can be limited during the training phase, reducing the interference of irrelevant slots on the current state judgment.

[0030] Based on the slot dependency graph, a state transition rule graph is further constructed. The state transition rule graph includes at least three types of constraints: business constraints, temporal constraints, and mutual exclusion constraints. Business constraints restrict the legal combinations of multiple slots within the same business domain; temporal constraints restrict the order of time-related and process-related slots; and mutual exclusion constraints restrict state combinations that cannot be simultaneously valid. The construction can be completed using a "business rule analysis—historical dialogue statistics—manual verification and confirmation" approach. This involves first providing initial rules based on the business process, then statistically analyzing high-frequency state transition patterns from the training corpus, and finally manually correcting conflicting rules. The state transition rule graph obtained in this way not only provides legality supervision during the training phase but also enables constraint filtering of candidate state update results during the inference phase.

[0031] Regarding the configuration of honorific sensitive slots, this implementation does not treat all slots as honorific sensitive. Instead, it only marks slots that affect title generation, politeness level selection, or service script template switching. For example, "예약자성별" can be used as an honorific sensitive slot in state modeling. Its role is that when the dialogue system needs to output confirmation statements, recommendation statements, or service feedback later, it can combine this slot with the current honorific level to determine a more appropriate Korean expression. The purpose of this setting is to ensure that the state tracking results not only reflect business semantics but also provide stable contextual conditions for target language generation in translation-assisted scenarios.

[0032] Step two involves first organizing the Korean multi-turn dialogue training corpus according to the dialogue session dimension, and then arranging the dialogues within each session in the actual interaction order to form a round-level sample sequence. Each round sample retains the original Korean text of the current round, as well as the dialogue state before and after the current round, so that the model training object is no longer an isolated sentence, but a round record with a clear state evolution path. In specific construction, complete dialogues can be extracted from manually annotated corpora or business logs, and then combined with the confirmed state of the previous round to generate the dialogue state before the update of the current round, and the updated dialogue state can be generated based on the actual semantic update result of the current round. In this way, the samples naturally carry the link information of "where the state comes from, what kind of update it goes to," which can avoid the problem of ambiguity in the supervision signal caused by directly learning the final state from a single round of input.

[0033] After generating the pre-update and post-update dialogue states, a slot-by-slot comparison is performed on all slots to form a slot update differential mask and a slot update type label. Specifically, the two states are first aligned on a unified slot set, and then each slot is checked individually to determine if it has changed. If a slot's value is empty in the pre-update dialogue state but not empty in the post-update state, the slot is considered an addition; if it's not empty before the update but empty after, it's considered a deletion; if both are not empty but have different values, it's considered a replacement; if both have the same value, it's considered a retention. Simultaneously, whenever a slot has different values, or one is empty while the other is not, the differential mask corresponding to that slot is marked as an update state; otherwise, it's marked as a retention state. This "unified alignment first, then slot-by-slot judgment" construction method ensures consistent supervision across different business domains and rounds, allowing the model to directly learn the state changes themselves, rather than just learning static state content.

[0034] In this implementation, the key round identifier is not arbitrarily assigned manually, but derived from the number of slots that are actually added, deleted, or replaced in the current round. Specifically, when at least one slot is determined to be added, deleted, or replaced, the round can be marked as a key round. When it is necessary to further differentiate the update intensity, the number of changed slots can also be used as the basis for calculating the key level and mapped to key round identifiers of different levels. The advantages of this approach are: on the one hand, it can distinguish the rounds that truly trigger state changes from a large number of unupdated rounds; on the other hand, it can increase the supervision intensity of these rounds during training, allowing the model to focus on learning "which expressions lead to state changes" and "which slots the state changes specifically occur in," thereby reducing the dilution effect of a large number of state-maintaining samples on updated samples in multi-round dialogues.

[0035] Furthermore, the seven-tuple samples, slot update differential masks, slot update type labels, and key round identifiers formed in step two are not isolated from each other, but rather work together to serve the subsequent training process. Specifically, the seven-tuple samples preserve the complete round context and state evolution basis; the differential mask directly indicates which slots require priority attention in the current round; the update type label distinguishes between three different update mechanisms—addition, deletion, and replacement—preventing the model from treating all state changes indiscriminately; and the key round identifier provides stronger supervision for rounds where real state changes occur during sample extraction, loss weighting, or hard sample selection. By constructing these intermediate results simultaneously, subsequent model training can be transformed from "directly predicting the final state" to a hierarchical learning process of "first identifying changes, then determining the type of change, and finally learning the result of the change," thereby improving the stability and interpretability of the training process. Step three: Upon receiving the original Korean text for the current round, the character sequence, word order, and original word segmentation boundaries are preserved. No particle stripping, word ending truncation, or replacement rewriting is performed. Instead, morphological tags, honorific level tags, and particle anchor point location tags are added to the original text. This approach is adopted because slot cues in Korean dialogues often appear alongside particles, word endings, and preceding / following collocations. Directly disrupting the original sentence structure during preprocessing can easily cause slot evidence such as location, time, and object to become detached from the original context, thus affecting subsequent candidate evidence location and state update judgments.

[0036] In the light morphological analysis stage, stemming, part-of-speech, or morphological category information is added to each word in the current round of original Korean text, without changing the word's position in the original sentence or disassembling it into new replacement text. The purpose of this approach is to allow the model to preserve the word order, connections, and local collocations of the original Korean expression while obtaining sufficient basic structural information to support subsequent state tracking. Compared to the approach of full morphological segmentation followed by sentence reconstruction, this implementation emphasizes "additional annotation" rather than "replacement rewriting," providing a unified original text coordinate system for subsequent slot anchor point location, evidence fragment extraction, and restored text fusion.

[0037] During the particle anchoring stage, priority is given to marking the positions of particles such as “이 / 가”, “을 / 를”, “은 / 는”, “에”, “에서”, and “(으)로”, which are highly relevant to slot location in Korean dialogue. Specifically, without altering the word segmentation boundaries of the original Korean text, the position of the word or word block containing the aforementioned particles in the current round of the original Korean text is identified, and this position is recorded as a particle anchor position label. The particle anchor position label does not output new text independently but is used as boundary limiting information for subsequent candidate evidence. For example, when a particle anchor corresponds to a location expression, candidate evidence for subsequent location slots is preferentially extracted from the vicinity of that particle; when a particle anchor corresponds to a time expression or an object expression, the range of candidate evidence for the corresponding slot is limited to local segments that form a stable pairing with that particle. This approach reduces noise interference caused by the model's unconstrained search for evidence across the entire sentence.

[0038] In the honorific recognition stage, honorific level labels are mainly generated by recognizing Korean terminator endings and main honorific markers. Specifically, the final forms corresponding to formal, polite, and informal forms can be identified from the end of the original Korean text or predicate-related positions in the current round, and the politeness level of the expression in the current round is determined by combining this with the main honorific marker. The generated honorific level labels do not directly replace the original text or change the original sentence content, but rather serve as auxiliary constraints for subsequent state update reliability assessment. Especially in the state tracking of request strength-related slots or title-related slots, the same business semantics may correspond to different update priorities or different expressive intentions under different honorific levels. Therefore, introducing honorific levels as additional labels helps improve the stability of these slot update judgments. Step four involves the following steps: First, the range of slots suspected of undergoing state changes in the current round is determined based on the slot update differential mask. Then, the original Korean text of the current round is checked for any missing explicit semantic components corresponding to these suspected update slots. Only when a suspected update slot exists in the current round, and the required subject, object, or key supplementary fragments for that slot are not fully present in the original Korean text of the current round, does the omission recovery scoring stage begin. This "first determine suspected updates, then determine if missing" approach avoids invalid recovery for rounds where no state change has occurred, thereby reducing the interference of recovery noise on subsequent state tracking.

[0039] During the triggering phase, the subject omission indicator, object omission indicator, and partial omission indicator are calculated separately. The subject omission indicator is determined based on whether the current round contains an action or judgment expression but lacks a clear agent; the object omission indicator is determined based on whether the current round contains a transitive expression but lacks a corresponding patient; and the partial omission indicator is determined based on whether the current round only contains a short response, a transitional connection, a negative confirmation, or a partial correction expression. Subsequently, the above three types of indicators are weighted and summed according to preset weights to obtain an omission trigger score. When this score is not lower than a preset recovery threshold, it is determined that the current round needs to perform omission recovery. The preset weights and recovery threshold can be set according to the frequency of different omission types and the cost of misjudgment in the development corpus, so that the recovery trigger can cover high-frequency omission cases while avoiding over-recovery of rounds with complete expressions.

[0040] After determining that omission recovery is required, a candidate recovery set is constructed from the historical round information within the candidate window. Specifically, the two to four most recent rounds preceding the current round are used as candidate windows, prioritizing the extraction of noun phrases related to the suspected update slot, confirmed slot values, and dependent slot values ​​adjacent to the suspected update slot. The candidate range is limited to the two to four most recent rounds because omission completion in Korean multi-turn dialogues usually relies on the nearest neighbor context, and excessively long windows can easily introduce irrelevant information from earlier rounds. After the candidate recovery set is established, the matching degree between each candidate and the suspected update slot is scored by combining the original Korean text of the current round and the dialogue state before the update, and the target candidate with the highest score is selected to form the recovery text. This recovery text does not rewrite the entire sentence, but only fills in the semantically missing parts corresponding to the suspected update slot, thus ensuring that the recovery result always revolves around the state update needs of the current round. Step 5: First, based on the slot update differential mask obtained in Step 2, determine the set of slots where the current round's state may change. Then, organize the evidence slot by slot. The purpose of this setup is to refine the judgment of "whether to update" and "why update" from sentence-level processing to slot-level processing, avoiding mutual interference when multiple slots appear simultaneously. The evidence chain for each slot consists of at least three types of evidence: original text evidence fragments, restored evidence fragments, auxiliary word anchor evidence, honorific level evidence, and translation auxiliary evidence. This ensures that subsequent state write-backs do not rely solely on a single path output but are based on multi-source semantic support.

[0041] When constructing the original evidence fragment, the first step is to perform slot anchor point localization in the current round of original Korean text around the candidate slot. Specifically, this can be done by combining the allowed value set for the slot, the high-frequency expression templates related to the slot in the business domain, the keyword positions in the current round of text, and the aforementioned particle anchor point position labels, to jointly determine the candidate evidence interval for the slot in the current round of original Korean text. If there is an explicit expression directly related to the slot in the current round of original Korean text, the corresponding phrase, short phrase, or partial sentence segment is extracted as the original evidence fragment; if there is no complete explicit expression in the current round of original Korean text, but omission recovery has been triggered in step four, then the same slot is further located in the recovered text, and the supplemented partial content is extracted as the recovered evidence fragment. By retaining the original evidence and the recovered evidence separately, the subsequent model can utilize both the "user's original expression" and the "expression supplemented based on context," thereby reducing the risk of misjudgment caused by relying solely on a single path.

[0042] In this step, particle anchor evidence and honorific level evidence primarily serve a constraining and corrective role, rather than solely determining the slot update result. Specifically, particle anchor evidence consists of the particle position and type corresponding to the candidate slot, used to further limit the boundaries of the original text evidence fragment and the restored evidence fragment. For example, when the candidate slot belongs to the location, time, or object category, priority is given to retaining the local fragments adjacent to particles such as "에", "에서", "(으)로", and "을 / 를" as the scope of valid evidence, avoiding other irrelevant components in the whole sentence from mistakenly participating in the slot judgment. Honorific level evidence consists of the honorific level labels of the original Korean text in the current round, mainly used to limit the update credibility of slots related to request intensity or title. That is to say, when the candidate slot is related to the level of politeness, the address of the service recipient, or the tone of the request, the slot update should be judged in conjunction with the honorific level in the current round to determine whether it is consistent with the overall expression style, so as to reduce the situation of making incorrect updates based solely on the literal content.

[0043] In this implementation, translation-assisted evidence is not included in the evidence chain for each slot by default, but is conditionally introduced only when "the evidence for the current slot is insufficient but still has auxiliary value." Specifically, for each suspected update slot, the uncertainty of that slot is calculated, and candidate translation-assisted texts and basic confidence scores corresponding to that slot are obtained from parallel corpus sources, target domain slot mapping dictionary sources, and translation model sources, respectively. Then, the basic confidence scores provided by each translation-assisted source for that slot are weighted and summed according to their respective historical accuracy weights to obtain the slot-level fusion confidence score; where the sum of all fusion weights is one. Only when the uncertainty of the corresponding suspected update slot is higher than a preset uncertainty threshold, and the slot-level fusion confidence score is not lower than a preset assistance threshold, is the translation-assisted evidence allowed to be included in the evidence chain for that slot and participate in subsequent consistency evaluation. This method of independent calculation per slot and conditionally controlled introduction avoids unnecessary interference with the original Korean semantics when the evidence is sufficient, and also avoids low-quality translation sources misleading state updates.

[0044] After forming a complete slot-level evidence chain, a slot evidence consistency assessment is further performed. Specifically, at least three types of valid evidence from the original text evidence fragments, restored evidence fragments, particle anchor evidence, honorific level evidence, and translation-assisted evidence are jointly evaluated to assess whether they collectively support the same slot update conclusion. If multiple types of evidence all point to the same slot being updated to the same target value, or collectively support that the slot should not be updated, then the slot is judged to have a high consistency result. If different pieces of evidence have obvious conflicts regarding the direction of slot update, such as original text evidence supporting replacement, restored evidence supporting preservation, and translation-assisted evidence supporting another value, then the consistency result of the slot is reduced. Overall, the more valid evidence there is and the more consistent the support direction of various types of evidence for the same slot update conclusion, the higher the evidence consistency result of the corresponding slot; conversely, the lower the consistency result. Through this assessment method based on the consistency of multiple types of evidence, step five can provide a more robust decision basis for subsequent state write-back than simple model scoring.

[0045] Step six, in specific processing, firstly, the original Korean text of the current round is used as the basic input, then the restored text is used as local supplementary input, morphological labels, honorific level labels, and particle anchor point position labels are used as structural auxiliary input, the dialogue state before the update and the historical round information within the historical context window are used as cross-round state reference input, the slot update differential mask is used as a hint input for the suspected update range, and the slot-level evidence chain is used as supporting input for candidate updates. Through this joint input method, when processing the current round, the model no longer relies solely on the surface text of a single sentence, but can establish a unified correlation between the original expression, the historical state continuity relationship, and multi-source evidence support, thereby improving its ability to identify real state changes.

[0046] During internal model processing, the aforementioned joint input is first used to form the slot-level representation of the current round. Specifically, the original Korean text and the recovered text jointly provide the semantic content of the current round; morphological labels, honorific level labels, and particle anchor position labels are used to limit the semantic interpretation boundaries; the pre-update dialogue state and historical context are used to provide the state continuation background; the slot update differential mask is used to highlight the slot positions in the current round that are more likely to change; and the slot-level evidence chain is used to provide additional support for the candidate update conclusions. Based on this, the model outputs six types of results: first, slot existence results, used to determine whether each slot in the current round is activated and needs to participate in the update; second, slot value prediction results, used to provide candidate update values; third, slot update gating results, used to determine whether candidate slots have entered the actual write-back process; fourth, state transition legality results, used to determine whether the current candidate state satisfies the state transition rule graph constraints; fifth, recovery candidate selection results, used to reflect whether the omitted recovery path is reliable; and sixth, slot evidence consistency results, used to characterize whether multi-source evidence jointly supports the update conclusion of the slot. By outputting these results simultaneously, model training and subsequent state write-back are no longer driven by a single output, but are supported by multiple intermediate judgments.

[0047] In this implementation, the selective write-back in step seven is used to transform the intermediate results output in step six into a controllable state update action. Specifically, it doesn't simply overwrite the original slot value in the pre-update dialogue state as soon as the model provides a new slot value prediction result. Instead, it checks whether the slot meets three conditions sequentially: First, the slot update gating result indicates that the slot indeed has the necessity to enter the update process; second, the state transition legality result indicates that the slot's transition from the pre-update state to the candidate updated state does not violate the business constraints, temporal constraints, and mutual exclusion constraints in the state transition rule diagram; third, the slot evidence consistency result reaches a preset evidence consistency threshold, indicating that the update conclusion has obtained sufficient and consistent evidence support. Only when all three conditions are met simultaneously is it allowed to write the candidate slot value back to the updated dialogue state; otherwise, the original state value in the pre-update dialogue state remains unchanged. The purpose of this controlled write-back method is to avoid incorrect state overwriting due to single prediction fluctuations, local recovery errors, or translation aid bias.

[0048] For slots that do not meet the write-back conditions, this implementation does not simply discard the results, but retains them as anomalous samples and proceeds to subsequent analysis. Specifically, when a slot is blocked by update gating, deemed illegal by legality verification, or has insufficient evidence consistency, the corresponding sample will record the original Korean text of the current round, the restored text, the dialogue state before the update, candidate update values, various intermediate results, and the reason for not writing back. The advantage of this approach is that the system can clearly distinguish between two different problems: "the model did not recognize the update" and "the model recognized the update but failed the verification." This provides directly usable sample evidence for subsequent difficult sample re-training and targeted retraining, thus forming a closed loop from online judgment to offline optimization.

[0049] In step eight, the purpose of setting the joint loss function is not simply to superimpose multiple training objectives, but to enable the model to simultaneously learn six categories of capabilities: which slots need attention in the current round, what values ​​these slots should be updated to, whether the update is legal, whether the recovery path is correct, whether the differential prediction is accurate, and whether the multi-source evidence is consistent. Specifically, the slot existence classification loss constrains the model to identify which slots are activated or participate in the update in the current round; the slot value prediction loss constrains the model to output the correct slot value; the state transition legality discrimination loss constrains the model to learn business rules and state transition boundaries; the recovery candidate selection loss constrains the model to select an appropriate recovery path when omissions occur; the slot update differential consistency loss constrains the model to correctly identify the boundary between update and maintenance; and the slot evidence consistency loss constrains the update results output by the model to obtain the common support of multi-source evidence. The above six losses are multiplied by their corresponding weight coefficients and then weighted and summed, so that the training process focuses on the correctness of the results, as well as the update path, rule compliance, and evidence sufficiency.

[0050] In this implementation, the weight coefficients are not arbitrarily set, but rather used to balance the impact of different sub-tasks on the overall training objective. Specifically, the first and second weight coefficients correspond to the slot existence classification loss and slot value prediction loss, respectively, and have relatively high values ​​because the fundamental goal of state tracking remains correctly identifying slots and correctly outputting slot values. The third and fifth weight coefficients correspond to the state transition legality discrimination loss and slot update differential consistency loss, respectively, and have relatively moderate values ​​because these two parts are mainly used to constrain update boundaries and migration paths. The fourth weight coefficient corresponds to the recovery candidate selection loss, with a value range of 0.3 to 1.0, to provide sufficient supervision in rounds that trigger omitting recovery, but to avoid exerting too strong an influence on ordinary rounds that do not trigger recovery. The sixth weight coefficient corresponds to the slot evidence consistency loss, with a value range of 0.2 to 1.0, used to ensure that the multi-source evidence support capability plays a stable constraining role in the overall training. By presetting the above ranges, the contribution of each task can be adjusted under different training data scales and different business domain conditions, thereby ensuring that model training has a repeatable parameter control basis.

[0051] In this implementation, the establishment of a hard sample library is used to separately manage and repeatedly reinforce the samples that are most prone to errors during training and have a significant impact on state tracking results. Specifically, when any of the following situations occur, the corresponding sample is added to the hard sample library: the slot value prediction confidence is lower than a preset threshold, indicating that the model's judgment of slot values ​​is unstable; the state transition legality result is characterized as an illegal transition, indicating that the candidate update violates the rule constraints; the highest score of the recovery candidate is lower than a preset recovery threshold, indicating that the omitted recovery path is unreliable; the slot update differential mask prediction result is inconsistent with the annotation result, indicating that the model has an error in identifying the update boundary; the slot evidence consistency result is lower than a preset evidence threshold, indicating that the candidate update lacks sufficient support. At the same time, at least one of the following samples—misjudging honorific level, slot extraction error samples caused by particle ambiguity, failed omission recovery samples, illegal state transition samples, differential mask prediction error samples, and samples with insufficient evidence consistency—is given a hard sample weight of 1.5 to 3.0, and participates in retraining in subsequent training rounds with a higher sampling weight than ordinary samples. In this way, the model can repeatedly learn the key error types that are most likely to affect the stability of state updates, thereby gradually improving its adaptability to complex Korean dialogue scenarios.

[0052] 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 training method for a Korean dialogue state tracking model, characterized in that; Includes the following steps: Step 1: Establish a set of target business domains, a native Korean slot system, a slot dependency graph, and a state transition rule graph for Korean multi-turn dialogue in translation assistance scenarios. The native Korean slot system includes a set of slots corresponding to each target business domain, a set of allowed values ​​corresponding to each slot, and a set of sensitive slots associated with honorific expressions or native Korean service scenarios. The state transition rule graph includes business constraint relationships, temporal constraint relationships, and mutual exclusion constraint relationships. Step 2: Organize the Korean multi-turn dialogue training corpus into a turn-level sample sequence, and construct each turn training sample as a seven-tuple sample including dialogue identifier, turn identifier, speaker identifier, original Korean text of the current turn, dialogue state before update, dialogue state after update, and key turn identifier; Based on the slot-by-slot comparison results between the dialogue state before update and the dialogue state after update, generate a slot update differential mask and a slot update type label. Specifically, when the slot value of a certain slot is different in the dialogue state before update and the dialogue state after update, or one is empty while the other is not empty, the differential mask corresponding to that slot is determined to be the updated state; otherwise, it is determined to be the maintain state. Step 3: Perform Korean preprocessing on the original Korean text for the current round. Korean preprocessing includes at least light morphological analysis, honorific recognition, and particle anchor labeling. The original Korean text character sequence and word order are preserved, and only morphological labels, honorific level labels, and particle anchor position labels are added. Particle separation and replacement and word ending cutting and replacement are not performed. Step 4: Based on the slot update differential mask, the original Korean text of the current round, the dialogue state before the update, and the historical round information in the candidate window, perform the omission recovery trigger judgment; when the omission trigger score after weighting the subject omission indicator, object omission indicator, and partial omission indicator according to their respective weights is not lower than the preset recovery threshold, omission recovery is triggered; generate a candidate recovery set from the noun phrases, confirmed slot values, and dependent slot values ​​adjacent to the suspected update slot in the most recent two to four historical rounds, and select the target candidate to form the recovery text, whereby the recovery text is only used to supplement the semantic missing part corresponding to the suspected update slot; Step 5: Construct a slot-level evidence chain for each suspected update slot. The slot-level evidence chain shall include at least three of the following: original text evidence fragments from the original Korean text, restored evidence fragments from the restored text, particle anchor evidence, honorific level evidence, and translation auxiliary evidence. Among them, the translation auxiliary evidence shall come from at least two of the following: parallel corpus sources, target domain slot mapping dictionary sources, and translation model sources. The translation auxiliary evidence shall be included in the slot-level evidence chain only when the uncertainty of the suspected update slot is higher than the preset uncertainty threshold and the corresponding slot-level fusion confidence is not lower than the preset auxiliary threshold. Step 6: Input the original Korean text of the current round, the restored text, the morphological label, the honorific level label, the particle anchor position label, the dialogue state before the update, the historical round information in the historical context window, the slot update differential mask, and the slot-level evidence chain into the Korean dialogue state tracking model, and output the slot existence result, slot value prediction result, slot update gating result, state transition legality result, restoration candidate selection result, and slot evidence consistency result. Step 7: Based on the slot update gating results, state transition legality results, and slot evidence consistency results, perform selective write-back on the candidate updated dialogue states. Specifically, only slots that simultaneously meet the update gating conditions, state transition rule graph constraints, and evidence consistency threshold conditions are written back. Slots that do not meet any of the conditions retain the original state values ​​in the dialogue states before the update and are recorded as abnormal samples. Step 8: Construct a joint loss function based on slot existence results, slot value prediction results, state transition legality results, recovery candidate selection results, slot update differential mask, and slot evidence consistency results. Then, use the joint loss function to update the parameters of the Korean dialogue state tracking model to obtain the trained Korean dialogue state tracking model. The joint loss function consists of the following weighted sums: slot existence classification loss multiplied by the first weight coefficient, slot value prediction loss multiplied by the second weight coefficient, state transition legality judgment loss multiplied by the third weight coefficient, recovery candidate selection loss multiplied by the fourth weight coefficient, slot update differential consistency loss multiplied by the fifth weight coefficient, and slot evidence consistency loss multiplied by the sixth weight coefficient. Specifically, when the confidence level of slot prediction is lower than the preset threshold, the legality result of state transition is characterized as illegal transition, the highest score of recovery candidate is lower than the preset recovery threshold, the prediction result of slot update differential mask is inconsistent with the labeling result, or the consistency result of slot evidence is lower than the preset evidence threshold, the corresponding training samples are written into the hard sample library and participate in retraining in subsequent training rounds with a higher sampling weight than ordinary samples.

2. The training method for a Korean dialogue state tracking model according to claim 1, characterized in that: The target business domain set includes at least three of the following: restaurant reservations, travel navigation, movie ticket reservations, express delivery tracking, and utility bill payment. Among them, the slot set of the restaurant reservation business domain includes at least three of the following: "음식종류", "맵기수준", "음식가격대", "요청사항", and "예약자성별". Among them, the slot corresponding to "예약자" is configured as a polite language sensitive slot, used to establish a mapping relationship with changes in Korean titles, changes in politeness levels, or changes in address templates.

3. The training method for a Korean dialogue state tracking model according to claim 1, characterized in that: The slot update type label includes at least three categories: addition, deletion, replacement, and retention. Among them, the addition type corresponds to slots that were empty in the previous dialogue state and were not empty in the new dialogue state; the deletion type corresponds to slots that were not empty in the previous dialogue state and were empty in the new dialogue state; the replacement type corresponds to slots that were not empty in both the previous and new dialogue states but had different slot values; and the retention type corresponds to slots that had the same slot value in both the previous and new dialogue states. Furthermore, the key round identifier is derived from the number of slots that were added, deleted, or replaced, and is used to improve the sample supervision strength for actual slot update rounds.

4. The training method for a Korean dialogue state tracking model according to claim 1, characterized in that: Particle anchor position labels are obtained by marking the position of at least one particle among "이 / 가", "을 / 를", "은 / 는", "에", "에서", and "(으)로" without changing the word segmentation boundaries of the original Korean text; Honorific level labels are obtained by identifying Korean terminology endings and main honorific markers; Furthermore, the particle anchor location tag is used to limit the candidate evidence boundary of the location slot, time slot, or object slot, and the honorific level tag is used to limit the update credibility of the request strength related slot or title related slot.

5. The training method for a Korean dialogue state tracking model according to claim 1, characterized in that: The original text evidence fragments in the slot-level evidence chain are obtained by performing slot anchor point positioning on the original Korean text of the current round. The recovered evidence fragments are obtained by performing same slot positioning on the recovered text. The particle anchor point evidence consists of the particle position and particle type corresponding to the candidate slot. The honorific level evidence consists of the honorific level label of the original Korean text of the current round. The translation auxiliary evidence consists of the translation auxiliary text corresponding to the candidate slot and its source confidence. Furthermore, the consistency result of slot evidence is obtained by conducting consistency assessments on at least three types of evidence, including original text evidence fragments, restored evidence fragments, particle anchor evidence, honorific level evidence, and translation auxiliary evidence. The more valid evidence participating in the consistency assessment and the more consistent the support direction of the various types of evidence for the update conclusion of the same slot, the higher the consistency result of the evidence for the corresponding slot.

6. The training method and system for a Korean dialogue state tracking model according to claim 1, characterized in that: For each suspected update slot, calculate the slot-level fusion confidence score separately; The slot-level fusion confidence is obtained by weighting and summing the basic confidence provided by each translation auxiliary source for the suspected updated slot according to the fusion weight corresponding to their respective historical accuracy. The sum of all the fusion weights is one. Furthermore, the corresponding translation auxiliary evidence is only allowed to participate in the evidence consistency assessment of the suspected update slot when the uncertainty of the corresponding suspected update slot is higher than the preset uncertainty threshold and the slot-level fusion confidence is not lower than the preset auxiliary threshold.

7. The training method and system for a Korean dialogue state tracking model according to claim 1, characterized in that: The first weight coefficient in the joint loss function ranges from 0.5 to 1.5, the second weight coefficient ranges from 0.8 to 2.0, the third weight coefficient ranges from 0.2 to 0.8, the fourth weight coefficient ranges from 0.3 to 1.0, the fifth weight coefficient ranges from 0.2 to 0.8, and the sixth weight coefficient ranges from 0.2 to 1.

0. Furthermore, a hard sample weight of 1.5 to 3.0 is applied to at least one of the following samples: samples with misjudged honorific level, samples with slot extraction errors caused by particle ambiguity, samples with failed omission recovery, samples with illegal state transitions, samples with errors in slot update differential mask prediction, and samples with slot evidence consistency results lower than a preset evidence threshold.

8. The training system for a Korean dialogue state tracking model according to claim 1, characterized in that, include: The slot system construction module is used to establish the target business domain set, the native Korean slot system, the slot dependency graph, and the state transition rule graph; The sample organization module is used to organize Korean multi-turn dialogue training corpus into seven-tuple samples; The differential label generation module is used to generate a slot update differential mask and a slot update type label based on the dialogue state before and after the update. The Korean preprocessing module is used to output morphological tags, honorific level tags, and particle anchor point position tags; The omission recovery module is used to perform omission recovery trigger judgment and generate recovery text; The evidence chain construction module is used to build a slot-level evidence chain for each suspected update slot, including original text evidence fragments, restored evidence fragments, auxiliary word anchor point evidence, honorific level evidence, and translation auxiliary evidence. The evidence consistency assessment module is used to output the evidence consistency results for each slot. The joint training module is used to output slot existence results, slot value prediction results, slot update gating results, state transition legality results, and recovery candidate selection results. The selective write-back module is used to perform selective write-back of candidate states based on the slot update gating result, the state transition legality result, and the slot evidence consistency result. The hard sample closed-loop module is used to write abnormal samples into the hard sample library and participate in subsequent retraining with a higher sampling weight than ordinary samples.

Images