Context anaphora resolution method and system based on dialogue structure perception

Abstract

The invention discloses a context anaphora resolution method and system based on dialogue structure perception, and relates to the field of natural language processing, and the method comprises the steps: obtaining multiple rounds of dialogue corpora, storing dialogue entities in a shared memory pool and a private memory pool respectively, clearly distinguishing entities with effective individual cognition and global common-known entities, and obtaining an entity with effective individual cognition and a global common-known entity; therefore, anaphora ambiguity caused by confusion of knowledge ranges of different speakers is eliminated; when anaphora resolution is carried out, through multi-layer structure signal fusion such as a significance recursion updating mechanism driven by a dialogue structure, semantic role consistency judgment and round correlation attenuation modeling, an interpretable cognitive compatibility scoring system is established, and an entity selection mechanism jointly constrained by multi-dimensional information is formed. Therefore, by introducing a dialogue structure perceived double-domain entity memory modeling mechanism, pronoun forepointing accuracy and robustness in complex scenes such as long dialogue, cross-speaker and multi-round interaction are effectively improved.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and in particular to a method and system for resolving contextual reference based on dialogue structure awareness. Background Technology

[0002] With the widespread application of conversational AI in customer service, virtual assistants, and multimodal interaction systems, one of the core challenges facing multi-turn dialogue semantic understanding is long-range referential resolution. In multi-turn interactions, participants repeatedly use pronouns such as "it," "he," and "there" to refer to previously mentioned entities or events. If the dialogue system cannot correctly identify the preceding objects of these pronouns, it will lead to misunderstandings. For example, in the sentence "A: Yesterday I went to the mall with Zhang San and Li Si. B: What did he buy?", determining who "he" refers to requires contextual analysis.

[0003] Current dialogue systems simplify tasks by transforming multi-turn dialogues into single-turn tasks and supplementing the preceding meaning of pronouns by expanding the text in each sentence, thereby flattening the contextual information. This can alleviate some pronoun ambiguity. However, they all focus on explicit text structure and ignore the information differences between dialogue participants. This makes it difficult for the system to capture implicit semantics or shared background information when facing situations where there are shared memories or private knowledge in the dialogue. As a result, semantic ambiguity or misinterpretation of reference is likely to occur, which affects the accuracy and coherence of the overall dialogue understanding. Summary of the Invention

[0004] This application provides a dialogue structure-aware contextual reference resolution method, system, storage medium, computer program product, and electronic device to at least solve the problem of pronoun reference confusion caused by the inability to effectively distinguish different contexts and knowledge sources in multi-turn dialogues in current related technologies.

[0005] In a first aspect, embodiments of this application provide a contextual reference resolution method based on dialogue structure awareness. The method includes: acquiring a multi-turn dialogue corpus, the multi-turn dialogue corpus including pronouns, speaker identities, and dialogue turn sequences; constructing a shared memory pool and multiple private memory pools based on the multi-turn dialogue corpus, wherein the private memory pools store private entities known only to the corresponding speaker, and the shared memory pools store shared entities known to all speakers; when a pronoun to be resolved is detected in the current turn of speech, extracting a set of candidate entities from the private memory pool and the shared memory pool of the current speaker; extracting entity salience, semantic role consistency, and turn correlation based on dialogue structure relationships, and calculating a cognitive compatibility score between each candidate entity and the pronoun to be resolved; the cognitive compatibility score is used to characterize the probability that a candidate entity becomes the referential target of the pronoun to be resolved; and selecting the candidate entity with the highest cognitive compatibility score as the referential object of the pronoun to be resolved.

[0006] Secondly, embodiments of this application provide a contextual reference resolution system based on dialogue structure awareness. The system includes: a corpus acquisition unit for acquiring multi-turn dialogue corpus, the multi-turn dialogue corpus including pronouns, speaker identities, and dialogue turn sequences; a memory pool construction unit for constructing a shared memory pool and multiple private memory pools based on the multi-turn dialogue corpus, wherein the private memory pools store private entities known only to the corresponding speaker, and the shared memory pools store shared entities known to all speakers; a candidate entity extraction unit for extracting a set of candidate entities from the private memory pool and the shared memory pool of the current speaker when a pronoun to be resolved is detected in the current turn of speech; a cognitive compatibility scoring unit for extracting entity salience, semantic role consistency, and turn correlation based on dialogue structure relationships, and calculating a cognitive compatibility score between each candidate entity and the pronoun to be resolved; the cognitive compatibility score is used to characterize the probability that a candidate entity becomes the referential target of the pronoun to be resolved; and a referential object selection unit for selecting the candidate entity with the highest score as the referential object of the pronoun to be resolved based on the cognitive compatibility score.

[0007] Thirdly, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the context reference resolution method based on dialogue structure awareness of any embodiment of the present application.

[0008] Fourthly, embodiments of this application provide a storage medium storing a computer program thereon, characterized in that, when the program is executed by a processor, it implements the steps of the context reference resolution method based on dialogue structure awareness of any embodiment of this application.

[0009] Fifthly, embodiments of this application provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the context reference resolution method based on dialogue structure awareness of any embodiment of this application.

[0010] The contextual reference resolution method and system based on dialogue structure awareness provided in this application can achieve at least the following technical effects: (1) By constructing a shared memory pool and a speaker's private memory pool in multi-turn dialogues, the system can dynamically limit the candidate entity set based on the visibility and cognitive scope of entities during referential resolution. This explicitly distinguishes between shared information and individual knowledge during dialogue reasoning, ensuring that the antecedent resolution of pronouns conforms to the actual cognitive context. When a pronoun appears, the system extracts candidate entities only from the current speaker's private and shared memory pools, fundamentally suppressing interference from irrelevant or invisible entities, thus making the selection of referential targets more constrained and reasonable. Therefore, based on shared and private memory modeling of the dialogue structure, when dealing with multi-turn interactions involving shared backgrounds or implicit knowledge, semantic consistency and referential coherence can be effectively maintained, significantly improving the accuracy of dialogue semantic understanding.

[0011] (2) After obtaining the candidate entity set, the cognitive compatibility score is calculated based on the entity saliency, semantic role consistency, and turn-based correlation extracted from the dialogue structure relationship. This scoring mechanism achieves unified quantification of semantics, pragmatics, and turn-based dynamics through the fusion of multi-dimensional features, so that the pronoun's referential judgment no longer depends on a single distance or text similarity, but comprehensively considers the entity's activity level and cognitive rationality in the dialogue process. Through the referential discrimination mechanism based on dialogue structure perception and cognitive compatibility calculation, the system can maintain stable referential judgment results in complex dialogue scenarios with multiple entity co-occurrence or topic switching, reducing semantic shifts and misjudgments, thereby achieving continuity and robustness of cross-turn semantic tracking.

[0012] This technical solution, by introducing dialogue structure awareness and explicit memory partitioning, transforms the pronoun reference problem from simple text similarity retrieval to entity-directed reasoning based on cognitive visibility and semantic structure. This makes the pronoun resolution process no longer dependent on flat context, but transforms it into a structure-driven, dynamic, interpretable, and inferable entity selection process, thereby systematically improving the accuracy and stability of multi-turn interactive semantic understanding. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A flowchart illustrating an example of a dialogue structure-aware context reference resolution method according to an embodiment of this application is shown. Figure 2A flowchart illustrating an example of constructing a memory pool based on a dialogue corpus according to an embodiment of this application is shown. Figure 3 A flowchart illustrating an example of extracting a set of candidate entities according to an embodiment of this application is shown. Figure 4 A flowchart illustrating an example of calculating a cognitive compatibility score according to an embodiment of this application is shown. Figure 5 This diagram illustrates an example of comparing the F1 scores of different models in different dialogue rounds. Figure 6 A block diagram of an example of a dialogue structure-aware context reference resolution system according to an embodiment of this application is shown. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] Currently, some experts and scholars have put forward some new insights into semantic understanding technology for multi-turn dialogues, but they mostly focus on the relationship between explicit text structure and graph structure, and lack modeling of the dynamics of speakers' private-shared memories and common knowledge.

[0017] More specifically, some studies have published dialogue understanding techniques based on sentence rewriting. Specifically, a readable expanded sentence is generated after each round of dialogue, using a pointer network to copy word fragments from historical statements. This method transforms multi-turn questions into single-turn question-answering tasks, improving parsing efficiency. However, because the expansion process relies on historical text, it cannot handle omitted information and speaker private memories, and rewriting errors can propagate to subsequent understanding. This study shows that over 70% of statements contain pronouns or omitted components, and rewriting errors significantly reduce understanding accuracy.

[0018] Some studies have also proposed attentional networks for core pronoun reference perception. For pronoun resolution in long-text dialogues, some research has transformed the pronoun resolution task into a graph node classification problem by constructing a graph structure that includes the speaker, turn of speech, and relationships between entities. Dual-channel attention networks propose simultaneously modeling linguistic and interactive information to address pronoun parsing errors caused by multiple jumps in reasoning. However, this graph structure processing has strong generality and struggles to capture implicit shared memories or the speaker's private cognition for specific dialogue scenarios; the connections between entity nodes also rely heavily on explicit textual cues, resulting in insufficient modeling of implicit knowledge.

[0019] Some studies have also proposed long context adaptation and dialogue rewriting techniques, which solve pronoun elimination in long documents through a four-step process (selecting candidate question-answer pairs, rewriting text using a pre-trained model, generating pseudo-labels, and fine-tuning the model), achieving a 3.61% F1 improvement. However, this method relies on repeated generation using a large model, resulting in high computational costs, poor real-time performance for multi-turn dialogues, and no consideration of speaker-specific knowledge.

[0020] Some studies have also proposed dialogue generation models based on shared and private memories. These models use shared memory to learn common features across different languages ​​while retaining private memory to capture the unique characteristics of each language. Experiments show that this method significantly improves model performance when training data is limited, demonstrating that distinguishing between shared and private memories can enhance the model's generalization ability in multi-domain, multi-agent scenarios. However, this framework has not yet been applied to pronoun resolution.

[0021] Some studies have also proposed shared memory datasets and long dialogue frameworks. To simulate conversations between real friends, dialogues from movie scripts were collected and annotated with speaker information, personal events, and shared memories between the two individuals. Research shows that incorporating shared memories makes the dialogue more coherent and that focusing solely on character settings or short-term events fails to reflect shared memories in close relationships, revealing the importance of shared memories in long-term conversations. However, how to utilize shared memories for pronoun deconstruction remains unexplored.

[0022] Some psycholinguistic experiments have found that when interpreting pronouns, comprehensionists prioritize candidate entities within shared knowledge. When a question involves the other party's private information, the respondent shifts their attention to the private area before incorporating the information into shared knowledge. Furthermore, when a speaker believes an entity is within the shared focus of attention, they directly use a pronoun without incorporation. This research suggests that human referential resolution relies on reasoning about the knowledge states of both parties, rather than simply on the surface structure of language.

[0023] It should be understood that the above description of the relevant technologies is intended only to help the public better understand the inventive spirit and motivation of this application, and is not intended to limit this application. Furthermore, the technical solutions described in the above-mentioned relevant technologies are not prior art, and may also be undisclosed technical solutions, such as those under research or in the laboratory stage.

[0024] The technical solutions in this application, including the collection, storage, use, processing, transmission, provision, and disclosure of users' personal information, comply with relevant laws and regulations and do not violate public order and good morals.

[0025] Figure 1 A flowchart illustrating an example of a dialogue structure-aware context reference resolution method according to an embodiment of this application is shown.

[0026] Regarding the execution entity of the method in this application, it can be any controller or processor with computing or processing capabilities, such as a multi-turn dialogue semantic understanding controller. By dividing the entity source into a shared memory pool and a private memory pool, the search space for candidate referential entities corresponding to pronouns is subject to visibility constraints. The system only performs pointing inference within the set of entities that the current speaker may actually perceive, thereby reducing semantic drift caused by cross-role and cross-knowledge boundaries. Under this mechanism, pronoun pointing no longer relies on simple literal proximity information, but can consciously distinguish the cognitive boundaries between different speakers. Therefore, in multi-turn long-range contexts, the stability and consistency of entity pointing are significantly enhanced.

[0027] In some examples, it may be integrated into an electronic device or terminal through software, hardware, or a combination of both, and the type of terminal or electronic device may be diverse.

[0028] like Figure 1 As shown, in step S110, a multi-turn dialogue corpus is obtained, which includes pronouns, speaker identities, and dialogue turn sequences.

[0029] In some implementations, multi-turn dialogues are constructed into a turn sequence in chronological order, with each turn's speech explicitly labeled with the speaker's identity, the content of the speech text, and the pronouns appearing within it. Simultaneously, language processing is performed on the speech text, including sentence segmentation, syntactic dependency analysis, semantic role identification, and entity extraction, to identify explicitly mentioned entities in sentences and their functional positions in the semantic structure, such as subject entities, object entities, modifying entities, or context entities. Thus, pronoun positions and syntactic roles (subject, object, prepositional phrase, etc.) are identified in each speech, and explicitly appearing entity names or descriptive noun phrases are extracted. This transforms the input text from unstructured natural language expression into a multi-turn dialogue corpus with a computable structure, characterizing the predicate-argument structure of "who is doing what to whom," providing a structured and parsable input basis for pronoun reference reasoning, and enabling entity state identification and dependency inference within a unified semantic reference framework.

[0030] In step S120, a shared memory pool and multiple private memory pools are constructed based on the multi-turn dialogue corpus. The private memory pool stores private entities known only to the corresponding speaker, while the shared memory pool stores shared entities known to all speakers.

[0031] Here, by distinguishing between shared knowledge and the speaker's private knowledge sources, the knowledge visibility boundaries between dialogue participants are explicitly differentiated, making the scope of candidate entities consistent with real pragmatic logic, so as to avoid erroneously including entities that do not have semantic accessibility in the inference scope.

[0032] In some implementations, a private memory pool is established for each speaker to store entities that are introduced only by that speaker and have not been cited, confirmed, or discussed by other parties. Simultaneously, a shared memory pool is established to store entity information that has been mentioned, confirmed, or repeated by two or more parties during the dialogue. Furthermore, when a private entity is accurately identified or cited by another speaker in a subsequent round, that entity is moved from the private memory pool to the shared memory pool to dynamically maintain the consistency of the overall cognitive state.

[0033] For example, when the first speaker introduces an entity that is only related to their personal experience in their speech, such as "I met with my mentor yesterday," the system stores "mentor" as a private entity in the speaker's private memory pool when parsing that round of speech. At this time, the second speaker cannot yet confirm the specific identity of this entity, so there is no sharing condition. However, when the second speaker actively asks in a subsequent round, "Did your mentor reply to your email later?" the system recognizes that the second speaker has clearly pointed to the same entity and has been able to form a semantic cognitive association with it. The entity then migrates from the first speaker's private memory pool to the shared memory pool, reflecting that the entity has now become common knowledge of the dialogue between the two speakers. Through this dynamic migration mechanism, the system can automatically correct the knowledge attribution status of entities during the continuous semantic evolution of multi-round interactions, ensuring that the memory pool structure remains consistent with the actual pragmatic sharing relationship. This avoids semantic deviations in subsequent referential inferences due to incorrect judgments of the entity's cognitive scope.

[0034] In step S130, when a pronoun to be resolved is detected in the current round of speech, a set of candidate entities is extracted from the private memory pool and the shared memory pool of the current speaker.

[0035] After a pronoun is detected, the scope of candidate entities is limited based on the speaker's visible knowledge, ensuring that the candidate selection process aligns with the actual cognitively available space. In some implementations, after a pronoun to be resolved is detected in the current round, candidate entities are extracted from the speaker's private and shared memory pools based on the speaker's identity. Entities that do not meet the semantic constraints are then excluded, taking into account the pronoun's grammatical category, semantic type, and whether it refers to a person or thing.

[0036] Furthermore, the contextual activity of entities can be assessed by combining factors such as the frequency of their appearance in recent rounds, their position in the most recently mentioned round, and their relevance to the current topic. Entities with significantly low activity or those that have degenerated in the context can be preferentially eliminated. Through candidate filtering and pre-screening, the candidate entity set maintains both necessary coverage and sufficient semantic validity, thereby providing an accurate and easily discriminative entity set.

[0037] In step S140, entity saliency, semantic role consistency and turn relevance are extracted based on the dialogue structure relationship. The cognitive compatibility score between each candidate entity and the pronoun to be resolved is calculated. The cognitive compatibility score is used to characterize the probability that the candidate entity becomes the referent of the pronoun to be resolved.

[0038] In some implementations, a multi-dimensional comprehensive index system is established based on factors such as the salience of candidate entities in recent rounds, their consistency with the semantic role in the current speech, the round adjacency between the entity and the current round, and the entity's visibility signal to the current speaker. This system performs a structured calculation of the matching degree between each candidate entity and the pronoun to be resolved. This calculation reflects whether the entity is in a naturally referential position, whether it reasonably assumes the current semantic function, whether it remains within the focus of the dialogue, and whether it is a semantic knowledge object reachable by the current speaker. Through this comprehensive scoring mechanism, the differences between candidate entities can be clearly distinguished, thus achieving robust differentiation even in the presence of multi-entity conflicts and semantic proximity competition.

[0039] In step S150, based on the cognitive compatibility score, the candidate entity with the highest score is selected as the referent of the pronoun to be resolved.

[0040] In some implementations, the entity with the highest cognitive compatibility score is selected from the candidate entity set as the referencing target, and after the referencing result is output, the memory attributes such as the active state and the most recent mention round of the corresponding entity are updated.

[0041] If the referencing result leads to the confirmation or reference of a private entity by another party, it can trigger the entity to migrate from the private memory pool to the shared memory pool, thereby reflecting the synchronous evolution of knowledge state in the continuous dialogue process. This avoids problems such as solidification of single-round decisions and lag in knowledge state, and enables the entire referencing inference process to have continuous correction capabilities and adaptability as the dialogue evolves, thereby improving the stability and overall accuracy of long-range multi-round referencing results.

[0042] Through the embodiments of this application, by integrating hierarchical modeling of the integrated memory pool with cognitive compatibility determination based on structure awareness, the system achieves a comprehensive understanding from information visibility constraints to semantic consistency assessment. This enables the system to accurately grasp pronoun referential relationships in dynamic, multi-participant dialogue environments, maintaining the coherence of semantic chains and the stability of contextual understanding. Compared to traditional methods that rely solely on text concatenation or unidirectional semantic similarity, this approach achieves deep, structured modeling of dialogue semantic understanding, significantly improving the accuracy of multi-turn dialogue referential resolution and overall comprehension performance.

[0043] In some examples of embodiments of this application, the multi-turn dialogue corpus has multi-dimensional semantic tags to simultaneously characterize the knowledge attributes of entities and the behavioral semantic structure of speaking turns. The multi-dimensional semantic tags include entity attribute tags and dialogue behavior tags. Entity attribute tags are used to define shared or private attributes of each entity in the dialogue, indicating whether the entity is information commonly known to all speakers. Furthermore, dialogue behavior tags are used to describe the semantic structural relationships of each speaking turn, including any of the following behavior types: asking a question, answering, confirming, or elaborating.

[0044] Here, entity attribute tags are used to indicate the visibility status of an entity among dialogue participants. For example, an entity can be labeled as a private attribute entity if it is only introduced by a speaker in their personal narrative and not cited by others. If the entity is explicitly restated or identified by other speakers in subsequent rounds, it will be relabeled as a shared attribute entity. Based on this, the tagging system makes the entity knowledge boundary no longer dependent on the explicit repetition logic of the text, but can be stably maintained through the "entity attribute tag - visibility boundary" method.

[0045] Simultaneously, at the semantic level, each round of speech is labeled with semantic behavior tags, including questioning, answering, confirming, or elaborating behaviors, to characterize the semantic role of the current round within the dialogue structure. For example, questioning typically points to the semantic emergence of the object being questioned, while confirming tends to focus on entities that were already clearly defined or factually constrained in the previous round, and elaborating tends to extend the expression along existing shared semantic lines. Through the labeling of these dialogue behavior types, the system can determine pronoun reference not only by relying on the text content itself but also by utilizing the dialogue intent constraints of the semantic behavior in that round to infer the type of entity that should be continued to be focused on or prioritized.

[0046] In some implementations, after acquiring and structuring multi-turn dialogue corpora, an automatic semantic parsing process is used to identify and extract labels from the entities and speaking behaviors. Specifically, based on entity recognition, dependency parsing, and entity occurrence relation judgment, the knowledge ownership attributes of each entity in the dialogue are automatically determined. If an entity appears only in the narration of a single speaker and is not repeated by another speaker, the system defaults to identifying it as a private attribute entity; if the entity is identified or cited by other speakers in subsequent turns, the system automatically classifies it as a shared attribute entity to maintain the dynamic consistency between entity attribute labels and real pragmatic visibility. In addition, by calling a machine learning model combined with syntactic and semantic behavior features, dialogue behavior judgment is performed on each turn of speech. If the current speech exhibits questioning characteristics, it is automatically labeled as a questioning behavior; if it semantically builds upon the information of the previous turn to affirm, deny, or modify, it is labeled as a confirmation behavior; and if the speech contains information extension and supplementation to an existing topic, it is labeled as an exposition behavior.

[0047] Through the embodiments of this application, a multi-dimensional semantic tagging system is constructed based on entity attribute tags and dialogue behavior tags. This system can form a two-layer constraint in the candidate entity screening stage and the referential inference stage, enabling the referential resolution process to explicitly utilize the differences in knowledge sources and pragmatic behavior in real conversations, thereby significantly improving the correctness, stability and cross-round consistency of referential recognition.

[0048] Figure 2 A flowchart illustrating an example of constructing a memory pool based on a dialogue corpus according to an embodiment of this application is shown.

[0049] like Figure 2 As shown, in step S210, entities are extracted from each round of dialogue corpus, and corresponding entity semantic vectors are generated based on the semantic features of the entities in the context.

[0050] In some implementations, entities in the current round of speech are automatically extracted based on word segmentation, entity recognition, dependency analysis, and semantic role labeling. These entities are then combined with their contextual features, syntactic position features, and association patterns to generate semantic vectors, which characterize the semantic content and semantic function relationships carried by the entities in a specific speaking context. This transforms unstructured natural language entity representations into standardized semantic vector representations, enabling the memory construction process to directly determine entity attributes and aggregate entities based on a unified semantic metric.

[0051] In step S220, the entity type of the corresponding entity is identified based on the entity attribute label. The entity type is either a shared entity known to all speakers or a private entity known only to the corresponding speaker.

[0052] Here, the visibility range of entities is distinguished from the knowledge source attribute dimension to ensure that the storage structure is consistent with the cognitive boundaries in the dialogue context. Specifically, based on the entity attribute label, it is identified whether the entity has been jointly mentioned or confirmed by multiple speakers in previous rounds. If it is an entity that all speakers can recognize and refer to, the system identifies it as a shared entity; if it only appears in a speaker's independent narration and is not pointed to or cited by other speakers, it is identified as a private entity of that speaker. When the entity attribute state changes, its attribute label can be automatically updated according to the dialogue content, so that the entity attribute label is consistent with the dynamic evolution of the dialogue, ensuring that the entity's type classification can accurately reflect the true semantic visibility range in the current dialogue.

[0053] In step S230, a private memory pool for the corresponding speaker is constructed based on the entity semantic vectors of each private entity and the corresponding dialogue behavior labels, and a shared memory pool is constructed based on the entity semantic vectors of each shared entity and the corresponding dialogue behavior labels.

[0054] The structured construction of the memory pool is driven by entity semantic vectors and dialogue behavior tags, enabling the separate management and independent use of private and shared knowledge in referential determination. Specifically, the system stores private entities in the private memory pool of the corresponding speaker based on their semantic vectors and behavioral participation patterns in multi-turn contexts, maintaining traceable entity memory within the individual's semantic scope. Simultaneously, shared entities are stored in the shared memory pool based on their semantic vectors and behavioral tags, reflecting their semantic reference function applicable to all speakers. The memory pool stores not only entity vectors but also information such as the entity's most recent turn, semantic relevance to the topic, and participation behavior patterns.

[0055] Therefore, by constructing private and shared memory pools differently, the system can accurately filter pronoun references in multi-turn dialogues based on knowledge sources and dialogue structure, thereby improving the controllability, interpretability, and accuracy of reference determination.

[0056] In some examples of embodiments of this application, for each private entity, the entity memory vector and salience score of the speaker to which the entity belongs are updated based on the entity semantic vector of the private entity and the corresponding dialogue behavior label.

[0057] Here, for each entity identified in the dialogue corpus, a private memory pool and a shared memory pool are constructed respectively. When the corpus sequence of the dialogue round arrives, the memory vector and salience of the entity are dynamically and recursively updated based on the entity semantic quantification representation and dialogue behavior labels, so as to accurately reflect the semantic activity and importance evolution of the entity in the dialogue process.

[0058] Specifically, in a given round, when a private entity appears or is mentioned, the system first performs a weighted fusion of the entity semantic vector corresponding to the current private entity in that round's corpus with the memory vector from the previous round. For example, the entity memory vector of a private entity is recursively updated using the following formula:

[0059] Equation (1) In the formula, Indicates the round number of the dialogue data. Indicates the speaker index, This represents private entities identified in the corpus. Entity index; and They represent the first Wheel and the first Speakers in turn-based dialogue material private entities The memory vector; For private entities In the The gating coefficient corresponding to the wheel, For private entities In the The semantic vector of the entity in the round corpus.

[0060] about The encoding details involve performing word-level segmentation on the text containing the private entity in the current round, and then semantically vectorizing the segmentation results using a pre-trained semantic model (such as a semantic encoder based on the Transformer structure). Subsequently, average pooling or Attention-weighted pooling strategies are used to aggregate all local word vectors corresponding to the entity in the current round, thereby obtaining the entity semantic embedding vector representing the true semantic expression of the private entity in the current round's corpus. This allows entity representations to closely reflect the semantic connotations of the entity within the context of the current dialogue, enhancing the local semantic accuracy during subsequent memory fusion.

[0061] In addition, through the gating coefficient By controlling the memory fusion ratio between the previous memory and the current semantic representation, a smooth transition between historical context memory and new semantic information can be achieved. Based on the gating memory mechanism, the cumulative drift of entity representation in long dialogue sequences can be effectively avoided.

[0062] The significance score of a private entity is recursively updated using the following formula: Equation (2) In the formula, The preset attenuation coefficient, For behavior-related activation coefficients, their values ​​are determined by the corresponding first... Dialogue behavior tags in turn-based dialogue data Decide; This is a vector transpose operation. It is the Sigmoid activation function. and They represent the first Wheel and the first Speakers in turn-based dialogue material private entities The significance score; Indicates the first The encoding vector of the turn-based dialogue corpus; For private entities, an indicator appears when the private entity... In the If a word is mentioned in the dialogue corpus, it is assigned a value of 1; otherwise, it is assigned a value of 0.

[0063] here, Used to characterize the The overall semantic expression features of turn-based dialogue corpora. Specifically, the first... The complete text of each round's speech is input into a sentence semantic encoder, which obtains the global semantic representation of that round through a multi-layer Transformer structure. The encoded output is then normalized and enhanced with an optional context window mechanism, ensuring that the semantic vector of the current round not only reflects the semantics of the single sentence in that round but also fully incorporates semantic background information related to the continuity of semantic topics from previous rounds. This forms a semantic embedding vector representing the global semantic state of that round. .

[0064] The saliency recursion mechanism can automatically model the linguistic patterns where proximity references are more important, and the semantic structure sensitivity can be enhanced by behavioral differentiation weighting, enabling the system to dynamically identify entities with greater referential potential in the current context as the dialogue progresses.

[0065] For each shared entity, the entity memory vector and salience score are updated based on the entity semantic vector and the corresponding dialogue behavior label.

[0066] The memory vector and saliency update method of shared entities are consistent with those of private entities, but they are not bound to a single speaker. Instead, they reflect the common semantic dependence of all participants on the concept, possessing cross-role semantic sharing attributes. The entity memory vector of a shared entity is recursively updated using the following formula:

[0067] Equation (3) In the formula, Represents shared entities identified in the corpus. Entity index, and They represent the first Wheel and the first Shared entities in round-robin dialogue corpus The memory vector, For shared entities In the The gating coefficient corresponding to the wheel, Represents shared entities In the The semantic vector of the entity in the round corpus.

[0068] The saliency score of shared entities is recursively updated using the following formula: Equation (4) In the formula, and They represent the first Wheel and the first Shared entities in round-robin dialogue corpus The significance score; For shared entities, an indicator appears when the shared entity... In the If a word is mentioned in the dialogue corpus, it is assigned a value of 1; otherwise, it is assigned a value of 0.

[0069] Through the above processing, it is ensured that when shared entities participate in cross-speaker semantic reasoning, their memory vectors always maintain the consistency of the common semantic space, so that the model has a stable semantic reference benchmark when performing cross-role and cross-rotation semantic reasoning.

[0070] In this embodiment, by constructing independent entity memory vector update mechanisms for private entities and shared entities respectively, and recursively updating them based on entity semantic vectors, gating coefficients and behavior-related activation coefficients in the dialogue rounds, entities can continuously maintain semantic expression capabilities with contextual continuity during multi-round dialogues.

[0071] Furthermore, by introducing a dual regulatory strategy of decaying salience scores over dialogue time and reactivating them based on behavior type, the importance of an entity naturally declines when it is not mentioned again, while its salience weight automatically increases when it reappears or is triggered by semantically significant behavior types. This better aligns with the dynamic shift in entity attention focus over time in real-world language interaction scenarios. Consequently, it ensures consistency, controllability, and interpretability in the semantic evolution of different types of entities in multi-speaker, multi-turn dialogue scenarios.

[0072] In some examples of embodiments of this application, different types of dialogue behavior labels are pre-configured with corresponding behavior-related activation coefficients, which are used to assign differentiated weights to entities involved in different dialogue behavior types when calculating saliency scores.

[0073] Specifically, different dialogue behaviors have different directions and magnitudes of influence on the importance of entities. For example, the "asking a question" behavior often guides the semantic focus to the entity being asked and forms a new point of interest, while the "confirming" behavior tends to strengthen the continuity between the semantic subject entity of the previous round and the current semantic core; the "answering" behavior emphasizes information supplementation and corresponding association, usually giving the previously asked entity a short-term significant boost; and the "elaborative" behavior tends to extend the entity dimension in the current context, and its boosting effect is relatively weaker than that of the "asking a question" and "answering" behaviors.

[0074] Therefore, by pre-setting different behavior-related activation coefficients for each behavior label and introducing them as dynamic adjustment quantities during the recursive calculation of saliency scores, the saliency scores not only reflect whether an entity appears, but also reflect the semantic function attributes and semantic influence strength of the appearing behavior, forming a fine-grained semantically sensitive dynamic adjustment mechanism, thereby significantly improving the expressive accuracy and contextual consistency of entity saliency evolution in multi-turn dialogues.

[0075] In some examples of embodiments of this application, the gating coefficients are dynamically generated through a learnable gating network. The gating coefficients are not fixed parameters or static thresholds, but are dynamically generated through the learnable gating network, thereby enabling the weights for updating entity memory to adaptively adjust according to the context state.

[0076] Specifically, the gating network constructs joint input features based on the encoded vector of the current round's dialogue corpus, the entity semantic vector of the target entity, and the dialogue behavior labels of the corresponding round. These features are then concatenated to form an input vector, which is linearly mapped using the learnable parameters of the gating network. Finally, a sigmoid activation function is used to output weight values, which determine the contribution ratio of the semantic information in the current round to entity memory updates. This is expressed by the following formula:

[0077] Equation (5) In the formula, This represents a vector concatenation operation. and These represent the learnable parameter matrix and bias term of the gated network, respectively.

[0078] Through the aforementioned dynamic gating mechanism, differentiated gating strategies can be learned for different contextual scenarios. In some scenarios, when the semantic information of a new round is strongly correlated with the entity, the gating coefficient automatically increases, causing the memory vector to converge quickly towards the current semantic direction; while when the semantic correlation is weak or the entity is not further strengthened, the gating coefficient tends to decrease to maintain the stability and continuity of historical memory. Compared with traditional fixed threshold or static function adjustment methods, the dynamic gating generation method can express the dynamic evolution characteristics of entity importance in multi-turn dialogues more finely, improving the semantic sensitivity and adaptability of entity state updates.

[0079] Figure 3 A flowchart illustrating an example of extracting a set of candidate entities according to an embodiment of this application is shown.

[0080] In this embodiment, the extraction process of the candidate entity set includes multiple parallel filtering steps to gradually narrow down the entity candidate space without disrupting the semantic continuity of the context, thereby improving the limitation and distinguishability of the referential resolution step.

[0081] like Figure 3 As shown, in step S310, time window filtering, saliency filtering, semantic type matching, and dialogue structure filtering are performed on entities in the private memory pool and the shared memory pool, respectively.

[0082] Here, time window filtering is used to select entities whose time frame is close to the current round's context. For example, the allowed effective semantic association range is determined based on the round index of the current round's dialogue corpus, and only entities that have appeared within the preset time window in the current round are retained. This reflects the short-term contextual validity of entities in natural dialogue, avoids entities that appear too early but are out of context from interfering with subsequent judgments, and improves the timeliness relevance of candidate entities.

[0083] Saliency filtering is used to retain entities with saliency scores higher than a preset threshold. For example, saliency filtering determines the retention of entities only by using the saliency scores obtained from private and shared entities during the aforementioned memory recursive update process. Since saliency scores represent the cognitive influence weight of an entity during the continuous evolution of the dialogue, saliency filtering can effectively exclude entities that exist semantically but lack contextual value, making the candidate set more focused on entities that currently have actual referential meaning.

[0084] Semantic type matching is used to retain entities of the same type based on the semantic category of the pronoun to be resolved. Specifically, semantic type matching filters entities based on the semantic category of the pronoun to be resolved to ensure that candidate entities belong to the same replaceable category as the pronoun in terms of semantic type. For example, personal pronouns are limited to entities of the person class, and infinitive pronouns are limited to entities of the item / place class, which helps to complete the semantic domain boundary shrinkage in advance at the entity candidate level and avoid the problem of candidate mis-diffusion caused by semantic domain confusion.

[0085] Dialogue structure filtering is used to limit the scope of entity associations based on the dialogue behavior labels of the current round. Specifically, dialogue structure filtering restricts based on the structural constraints corresponding to the dialogue behavior labels of the current round. For example, in "answer" type behaviors, entities associated with the semantic chain of the previous question are prioritized; in "explain" type behaviors, entities with strong salience from the previous round are prioritized; and in "confirm" type behaviors, shared entities are prioritized. Dialogue structure filtering reflects the pragmatic logical progression of human multi-turn dialogues, and can further constrain the entity candidate space at the dialogue behavior level, giving the candidate entity selection process a structured causal constraint semantic.

[0086] In step S320, the private entities and shared entities filtered by each filtering condition are merged and deduplicated to obtain a candidate entity set.

[0087] Specifically, the private and shared entities retained from each of the above filtering stages are merged and duplicates are removed to form the final candidate entity set. Since the aforementioned filtering processes all perform parallel contraction on semantic influence factors of different dimensions, and the constraint logic introduced in each filtering stage has independent semantic judgment capabilities, the final candidate entity set can significantly improve the semantic relevance and discrimination quality of candidate entities while ensuring coverage.

[0088] Figure 4 A flowchart illustrating an example of calculating a cognitive compatibility score according to an embodiment of this application is shown.

[0089] like Figure 4 As shown, in step S410, the semantic matching degree of the candidate entity's speech is determined.

[0090] It should be noted that semantic consistency refers to one of the necessary conditions for pronouns to be valid. Whether a candidate entity can be replaced by a pronoun in the current round depends primarily on the degree of proximity between the two in the semantic space. Therefore, the cosine similarity between the current round's statement encoding vector and the candidate entity's context semantic vector is used as the semantic matching degree of the speech.

[0091] Equation (6) In the formula, Indicates the index of the candidate entity. This indicates the round number of the current dialogue round. Indicates candidate entities With the current round of dialogue The semantic matching degree of the speech. This represents the encoded vector of the dialogue statement in the current round. Indicates candidate entities The semantic vector representation in the current round context.

[0092] In some implementations, if the statement is very short or the memory vector is sparse, it can be directed to... Adding temperature factor Perform soft scaling (e.g.) This ensures comparability across scenarios with different sentence lengths.

[0093] In step S420, the grammatical role matching degree of the candidate entities is determined.

[0094] It should be noted that the role of a pronoun in the syntactic structure (subject / object / prepositional object, etc.) is usually consistent with or similar to the role of the entity it refers to when it is mentioned. To portray syntactic role consistency.

[0095] Equation (7) In the formula, Indicates candidate entities With the current round The grammatical role matching degree reflects the consistency of the two in syntactic structure; This is the grammatical role vector of the current round of statements, representing the grammatical function of pronouns in the current sentence; This is the syntactic role vector corresponding to the candidate entity, representing the syntactic role features of the entity when it is mentioned.

[0096] Therefore, syntactic constraints are introduced into referential judgment to avoid incorrect forward referencing caused by semantic similarity alone.

[0097] In step S430, the round distance weight of the candidate entity is determined.

[0098] Equation (8) In the formula, Indicates candidate entities With the current round The round distance weight is used to represent the impact of the time when a candidate entity is mentioned on the probability of reference. This is the round decay coefficient, used to control the rate at which the matching score decays due to the round interval; Candidate entities The round number most recently mentioned; It is an exponential function used to convert the distance difference into a decay weight within the interval [0,1].

[0099] It should be noted that, considering the principle of the recency effect in human dialogue, i.e., mentions closer to the current round are more likely to be referred to earlier, in Equation (8), the round difference is continuously modeled through exponential decay, and the temporal constraints are supplemented by round distance weights, so that entities far from the context are naturally de-weighted. Regarding The round decay coefficient can still be adaptively set, with shorter sessions having a higher decay coefficient than longer sessions; when a topic restart flag exists, the decay coefficient should be appropriately increased. To accelerate cross-topic decay. Additionally, a truncation strategy can be employed when... The distance weight of each round is reset to 0 to prevent weak noise from the long-distance tail distribution.

[0100] In step S440, the saliency score of the candidate entity in the current round is obtained, and the speech semantic matching degree, grammatical role matching degree and candidate entity round distance are fused to calculate the cognitive compatibility score between the candidate entity and the pronoun to be resolved.

[0101] Equation (9) In the formula, Indicates candidate entities Cognitive compatibility score with the current round Indicates candidate entities The significance score in the current round, This is a weighting factor for the score. This is a bias term.

[0102] In equation (9), a weighted linear fusion method is used, followed by Sigmoid compression, to output... As a final compatibility measure. With bias Supervised training (using manually labeled leading references as positive samples and non-leading references as negative samples; loss can be binary cross-entropy or focus loss) allows for end-to-end optimization of the training process. By unifying attention (saliency) and constraints (semantics / syntacticity / temporality) into the same scoring framework, the output is a range-based, thresholdable probability metric, improving the accuracy, consistency, and interpretability of reference resolution.

[0103] This application establishes a multi-dimensional fusion-based cognitive compatibility scoring mechanism, enabling candidate entities to comprehensively consider their contextual semantic relevance, syntactic structural substitutability, turn-based temporality, and attentional weight within a unified probability space. This results in a more accurate forward inference that aligns with real-world human dialogue comprehension. The features across these dimensions both compensate for and counterbalance each other, preserving the flexibility and ambiguity of semantic expression while suppressing interference from invalid entities and cross-topic misreferences through structural information and attenuation mechanisms. This enhances the accuracy and robustness of pronoun referencing resolution in real-world multi-turn open dialogue scenarios.

[0104] To verify the effectiveness of the proposed method in this application, we constructed a dialogue dataset containing shared / private memory annotations and compared it with several baseline models. The experiments focused on evaluating the accuracy of referential resolution, the model's robustness to long contexts, and the performance variation under different memory weights.

[0105] Regarding data sources, we selected 200,000 multi-turn dialogues from public customer service conversations, movie scripts, and friend chat logs. We constructed pronoun referential markers, entity shared / private attributes, dialogue behavior, and sentiment tags through manual annotation. In some parts of the data, we borrowed methods from the SHARE dataset to extract shared memories from movie scripts.

[0106] Regarding the selection of baseline models, the following baseline models were chosen for comparison: Rewriting-based pronoun resolution: achieving pronoun parsing through text expansion; Graph-structured dual-channel attention network: using entity dependency graphs and combining language and interaction dual-channel attention as a representative graph theory model; Long-context Adaptation (LQCA): This solves the problem of long dependencies by rewriting text using a large number of generative models.

[0107] Regarding the model settings in the embodiments of this application, the memory vector dimension... The gating network and scoring function are trained using random initialization; weight parameters Selected by grid search on the validation set.

[0108] In terms of evaluation metrics selection, accuracy, recall, and F1 score are used to measure the effectiveness of the dereference resolution, and model latency is recorded to assess real-time performance.

[0109] Table 1 shows the index resolution performance of the Dual-Domain Dynamic Memory Model (D2DM) provided in this application embodiment and the baseline model on the overall dataset. It can be seen that the Dual-Domain Dynamic Memory Model significantly outperforms the baseline model in both accuracy and F1 score, while maintaining a lower response latency.

[0110] Table 1. Performance comparison of different models on the algebraic resolution task

[0111] Experimental results show that, compared with rewriting methods and graph structure models, the D2DM method can effectively capture long-term dependencies and implicit knowledge by introducing shared / private memory and cognitive compatibility scores, thereby improving the accuracy of reference resolution.

[0112] Figure 5 This diagram illustrates an example of the comparison results of F1 scores for referential resolution of different models at different dialogue rounds.

[0113] like Figure 5 As shown, the performance of both baseline models declines rapidly with increasing dialogue rounds, indicating their insufficient ability to handle long dialogue scenarios. In contrast, the D2DM method provided in this application maintains a significantly higher F1 score across all dialogue rounds, especially maintaining stable recognition capabilities in long dialogues exceeding five rounds. These experimental results demonstrate that by recursively adjusting and controlling the decay of entity semantic vectors and saliency weights through a dual-domain dynamic memory mechanism, the model can continuously retain key entity information related to long-range dependencies, avoiding the accumulation of errors caused by the continuous dilution of semantic cues in long dialogues. Therefore, compared to baseline methods that rely on rewriting or static graph structure representations, this application's embodiment exhibits stronger robustness and referential resolution capabilities under complex long dialogue conditions.

[0114] This paper addresses the challenge of handling implicit knowledge and shared memory in pronoun resolution within multi-turn dialogues, proposing a context-based pronoun resolution method based on dialogue structure awareness. By introducing the concepts of shared knowledge and private memory from psycholinguistics, a dual-domain dynamic memory model is constructed, and a cognitive compatibility scoring and knowledge-sharing gating mechanism are designed. The D2DM method significantly improves the accuracy and long-range robustness of pronoun resolution on multiple public and self-built datasets, and can adapt to long dialogues, complex emotions, and variable structures.

[0115] Future work can be carried out in the following aspects: (1) extending the dual-domain memory model to multi-party dialogue or group chat scenarios to handle complex shared knowledge among groups; (2) combining multimodal information, such as visual and audio cues, to further enhance memory representation; (3) exploring the feasibility of combining with large-scale language models, using this solution as a plug-in to improve the model's contextual understanding capabilities. We believe that by continuously exploring the unique cognitive features of dialogue scenarios, we can bring new breakthroughs to the field of natural language understanding.

[0116] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of combined actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Secondly, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application. In the above embodiments, the descriptions of each embodiment have their own emphasis; for parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0117] Figure 6 A block diagram of an example of a dialogue structure-aware context reference resolution system according to an embodiment of this application is shown.

[0118] like Figure 6 As shown, the contextual reference resolution system 600 based on dialogue structure awareness includes a corpus acquisition unit 610, a memory pool construction unit 620, a candidate entity extraction unit 630, a cognitive compatibility scoring unit 640, and a referential object selection unit 650.

[0119] The corpus acquisition unit 610 is used to acquire multi-turn dialogue corpus, which includes pronouns, speaker identities and dialogue turn sequences.

[0120] The memory pool construction unit 620 is used to construct a shared memory pool and multiple private memory pools based on the multi-turn dialogue corpus, wherein the private memory pool stores private entities known only to the corresponding speaker, and the shared memory pool stores shared entities known to all speakers.

[0121] The candidate entity extraction unit 630 is used to extract a set of candidate entities from the private memory pool and the shared memory pool of the current speaker when a pronoun to be resolved is detected in the current round of speech.

[0122] The cognitive compatibility scoring unit 640 is used to extract entity saliency, semantic role consistency and turn relevance based on dialogue structure relationship, and calculate the cognitive compatibility score between each candidate entity and the pronoun to be resolved; the cognitive compatibility score is used to characterize the probability that the candidate entity becomes the referent of the pronoun to be resolved.

[0123] The referential object selection unit 650 is used to select the candidate entity with the highest score as the referential object of the pronoun to be resolved based on the cognitive compatibility score.

[0124] In some embodiments, this application provides a non-volatile computer-readable storage medium storing one or more programs including execution instructions. The execution instructions can be read and executed by an electronic device (including but not limited to a computer, server, or network device) to perform the steps of any of the above-described context reference resolution methods based on dialogue structure awareness.

[0125] In some embodiments, this application also provides a computer program product, the computer program product including a computer program stored on a non-volatile computer-readable storage medium, the computer program including program instructions, which, when executed by a computer, cause the computer to perform the steps of any of the above-described dialogue structure-aware context reference resolution methods.

[0126] In some embodiments, this application also provides an electronic device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform steps of a dialogue structure-aware context reference resolution method.

[0127] The above-described product can perform the methods provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects for performing the methods. Technical details not described in detail in this embodiment can be found in the methods provided in the embodiments of this application.

[0128] The electronic devices in this application can exist in various forms, including but not limited to: mobile communication devices, ultra-mobile personal computer devices, portable entertainment devices, or other airborne electronic devices with data interaction functions.

[0129] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0130] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A contextual reference resolution method based on dialogue structure awareness, characterized in that, The method includes: Acquire multi-turn dialogue data, which includes pronouns, speaker identities, and dialogue turn sequences; Based on the multi-turn dialogue corpus, a shared memory pool and multiple private memory pools are constructed, wherein the private memory pools store private entities known only to the corresponding speaker, and the shared memory pools store shared entities known to all speakers; When a pronoun to be resolved is detected in the current round of speech, a set of candidate entities is extracted from the private memory pool and the shared memory pool of the current speaker. Based on the entity saliency, semantic role consistency and turn relevance extracted from the dialogue structure relationship, the cognitive compatibility score between each candidate entity and the pronoun to be resolved is calculated; the cognitive compatibility score is used to characterize the probability that the candidate entity becomes the referent of the pronoun to be resolved. Based on the cognitive compatibility score, the candidate entity with the highest score is selected as the referent of the pronoun to be resolved.

2. The method according to claim 1, characterized in that, The multi-turn dialogue corpus has multi-dimensional semantic tags, which include entity attribute tags and dialogue behavior tags; The entity attribute tags are used to define shared or private attributes of entities in a dialogue, indicating whether the entity is information known to all speakers. The dialogue behavior tags are used to describe the semantic structure relationship of each speaking round, including any of the following behavior types: asking a question, answering a question, confirming a question, or explaining a question.

3. The method according to claim 2, characterized in that, The construction of a shared memory pool and multiple private memory pools based on the multi-turn dialogue corpus includes: Extract entities from the dialogue corpus of each round, and generate corresponding entity semantic vectors based on the semantic features of the entities in the context; The entity type of the corresponding entity is identified based on the entity attribute label, wherein the entity type is a shared entity known to all speakers or a private entity known only to the corresponding speaker; A private memory pool for each speaker is constructed based on the entity semantic vectors of each private entity and the corresponding dialogue behavior labels, and a shared memory pool is constructed based on the entity semantic vectors of each shared entity and the corresponding dialogue behavior labels.

4. The method according to claim 3, wherein, The construction of a private memory pool for a speaker based on the entity semantic vectors of each private entity and the corresponding dialogue behavior labels, and the construction of a shared memory pool based on the entity semantic vectors of each shared entity and the corresponding dialogue behavior labels, include: For each private entity, update the entity memory vector and salience score of the speaker based on the entity semantic vector and corresponding dialogue behavior label of the private entity; The entity memory vector of the private entity is recursively updated using the following formula: ， In the formula, Indicates the dialogue corpus round number. Indicates the speaker index, This represents private entities identified in the corpus. Entity index; and They represent the first Wheel and First Speakers in turn-based dialogue material private entities The memory vector; For private entities In the The gating coefficient corresponding to the wheel, For private entities In the The semantic vector of the entity corresponding to the round corpus; The significance score of a private entity is recursively updated using the following formula: ， In the formula, The preset attenuation coefficient, For behavior-related activation coefficients, their values ​​are determined by the corresponding first... Dialogue behavior tags in turn-based dialogue data Decide; This is a vector transpose operation. It is the Sigmoid activation function. and They represent the first Wheel and First Speakers in turn-based dialogue material private entities The significance score; Indicates the first The encoding vector of the turn-based dialogue corpus; For private entities, an indicator appears when the private entity... In the If a word is mentioned in the dialogue corpus, it is assigned a value of 1; otherwise, it is assigned a value of 0. For each shared entity, the entity memory vector and salience score are updated based on the entity semantic vector and the corresponding dialogue behavior label of the shared entity; The entity memory vector of the shared entity is recursively updated using the following formula: ， In the formula, Represents shared entities identified in the corpus. Entity index, and They represent the first Wheel and First Shared entities in round-robin dialogue corpus The memory vector, For shared entities In the The gating coefficient corresponding to the wheel, Represents shared entities In the The semantic vector of the entity corresponding to the round corpus; The saliency score of shared entities is recursively updated using the following formula: ， In the formula, and They represent the first Wheel and First Shared entities in round-robin dialogue corpus The significance score; For shared entities, an indicator appears when the shared entity... In the If a word is mentioned in the dialogue corpus, it is assigned a value of 1; otherwise, it is assigned a value of 0.

5. The method according to claim 4, characterized in that, The gating coefficients are dynamically generated through a learnable gating network. ， In the formula, This represents a vector concatenation operation. and These represent the learnable parameter matrix and bias term of the gated network, respectively.

6. The method according to claim 4, characterized in that, Different types of dialogue behavior labels are pre-configured with corresponding behavior-related activation coefficients, which are used to assign differentiated weights to entities involved in different dialogue behavior types when calculating salience scores.

7. The method according to claim 4, characterized in that, The step of extracting a set of candidate entities from the current speaker's private memory pool and the shared memory pool includes: The entities in the private memory pool and the shared memory pool are respectively subjected to time window filtering, saliency filtering, semantic type matching and dialogue structure filtering; the time window filtering is used to filter entities that are close to the current round of context time; the saliency filtering is used to retain entities whose saliency scores are higher than a preset threshold; the semantic type matching is used to retain entities with the same type according to the semantic category of the pronoun to be resolved; and the dialogue structure filtering is used to limit the scope of entity association according to the dialogue behavior label of the current round. The private and shared entities filtered by each filter condition are merged and deduplicated to obtain a candidate entity set.

8. The method according to claim 4, characterized in that, The process involves extracting entity saliency, semantic role consistency, and turn-based relevance based on dialogue structure relationships, and calculating the cognitive compatibility score between each candidate entity and the pronoun to be resolved, including: Determine the semantic matching degree of the candidate entities' statements: ， In the formula, Indicates the index of the candidate entity. This indicates the round number of the current dialogue round. Indicates candidate entities With the current round of dialogue The semantic matching degree of the speech. This represents the encoded vector of the dialogue statement in the current round. Indicates candidate entities The semantic vector representation in the current round context; Determine the syntactic role matching degree of candidate entities: ， In the formula, Indicates candidate entities With the current round The grammatical role matching degree reflects the consistency of the two in syntactic structure; This is the grammatical role vector of the current round of statements, representing the grammatical function of pronouns in the current sentence; This is the syntactic role vector corresponding to the candidate entity, representing the syntactic role features of the entity when it is mentioned; Determine the round distance weights for candidate entities: ， In the formula, Indicates candidate entities With the current round The round distance weight is used to represent the impact of the time when a candidate entity is mentioned on the probability of reference. This is the round decay coefficient, used to control the rate at which the matching score decays due to the round interval; Candidate entities The round number most recently mentioned; It is an exponential function used to convert the distance difference into a decaying weight within the interval [0,1]. Obtain the saliency score of the candidate entities in the current round, and combine the speech semantic matching degree, the grammatical role matching degree, and the round distance of the candidate entities to calculate the cognitive compatibility score between the candidate entities and the pronouns to be resolved: , In the formula, Indicates candidate entities Cognitive compatibility score with the current round Indicates candidate entities The significance score in the current round, This is a weighting factor for the score. This is a bias term.

9. A contextual reference resolution system based on dialogue structure awareness, characterized in that, The system includes: The corpus acquisition unit is used to acquire multi-turn dialogue corpus, which includes pronouns, speaker identities, and dialogue turn sequences. The memory pool construction unit is used to construct a shared memory pool and multiple private memory pools based on the multi-turn dialogue corpus, wherein the private memory pools store private entities known only to the corresponding speaker, and the shared memory pools store shared entities known to all speakers; The candidate entity extraction unit is used to extract a set of candidate entities from the private memory pool and the shared memory pool of the current speaker when a pronoun to be resolved is detected in the current round of speech. The cognitive compatibility scoring unit is used to extract entity saliency, semantic role consistency and turn relevance based on dialogue structure relationships, and calculate the cognitive compatibility score between each candidate entity and the pronoun to be resolved; the cognitive compatibility score is used to characterize the probability that the candidate entity becomes the referent of the pronoun to be resolved; The referential object selection unit is used to select the candidate entity with the highest cognitive compatibility score as the referential object of the pronoun to be resolved.

Images