A long text cross-dialogue based reference resolution method

Abstract

This invention relates to the field of natural language processing technology and discloses a method for resolving referential inconsistencies across long texts in cross-dialogue contexts. The method includes: segmenting the long text into blocks and vectorizing it to form a queryable vector index set; constructing a hybrid label and semantic index by combining a business entity dictionary and historical weakly supervised signals to generate a candidate entity set; inputting the candidate entity set and block summaries into a dual-channel scoring model, where the first channel uses a zero-shot common-reference classifier to generate a probability score, and the second channel is calibrated based on business rules; weighting and fusing the two to generate the final matching score for candidate entity pairs; constructing a weighted entity graph and applying the maximum spanning tree algorithm to solve for the global preliminary referential chain structure; correcting and filling in conflicting entities appearing in the global preliminary referential chain structure by combining entity type, block summary, and context vector information to form a global referential chain set. This invention has the advantages of improving global consistency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a method for resolving referential issues in long texts through cross-dialogue. Background Technology

[0002] With the widespread application of artificial intelligence in customer service, cross-conversation and cross-channel text understanding and referential resolution have become key technologies. However, existing technologies still have many shortcomings in practical applications, which restrict the real-time performance and accuracy of intelligent customer service systems. The main problems include: bottlenecks in long text processing; in ultra-long conversations or multi-turn chat scenarios, the required GPU memory for the model increases dramatically, making it impossible for the system to efficiently process long text data; high dependence on fine-tuning; whenever the business line or application scenario changes, retraining is required, which is costly and difficult to deploy quickly; fragmentation of referential chains; the generated referential chains are easily fragmented, with a high entity association error rate, making it difficult to form a complete and consistent referential chain; insufficient real-time performance, resulting in significant processing delays for ultra-long text, affecting user experience and the response efficiency of intelligent customer service; channel fragmentation; it is not possible to effectively associate the same entity across multiple channels and conversations. Therefore, when facing long single conversations, multiple conversations, and multi-channel scenarios, existing technologies cannot achieve low latency, zero fine-tuning, and unified cross-channel processing capabilities while ensuring high-precision referential resolution. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method for resolving cross-dialogue references based on long texts, which has the advantages of improving global consistency and accuracy, and solves the problems mentioned in the background technology.

[0004] To achieve the aforementioned goals of improving global consistency and accuracy, this invention provides the following technical solution: a method for resolving pronoun references across long text dialogues, comprising the following steps: The text is divided into blocks based on length and semantic continuity. Entity references are retained, while non-entity sentences are compressed to generate block summaries and entity summary cards. These are then vectorized to form a queryable vector index set. Based on a queryable set of vector indexes, combined with a business entity dictionary and historical weak supervision signals, a hybrid index of tags and semantics is constructed. Cross-block, cross-session, and cross-channel retrieval and filtering are performed on each entity mention to generate a candidate entity set. The candidate entity set and the block summary are input into a dual-channel scoring model. The first channel uses a zero-shot common index classifier to generate a probability score, and the second channel is calibrated based on business rules. The two are weighted and fused to generate the final matching score of the candidate entity pair. Based on the final matching scores of candidate entity pairs, each candidate entity is treated as a node, and the final matching score between each pair of candidate entities is treated as the edge weight. A weighted entity graph is constructed, and the maximum spanning tree algorithm is applied to the weighted entity graph to solve the global preliminary referential chain structure. For conflicting entities appearing in the preliminary global referential chain structure, the system combines entity type, block summary and context vector information to correct and fill in the gaps, forming a globally consistent and callable global referential chain set.

[0005] Preferably, the process of forming a queryable vector index set is as follows: Semantic analysis algorithms are used to calculate the semantic similarity between sentences in the text, and the text is segmented into blocks in combination with text length constraints; High-ratio compression is performed on non-entity sentences within the block, while preserving key contextual information during the compression process, including core vocabulary in the sentence and semantic connections between sentences; Generate an intra-block summary for each text block, recording the original sentence position, keyword index, and sentence vector representation; For each entity in a text block, generate an entity summary card, including the entity name, location of occurrence, context window, and key tag information; The intra-block summary and entity summary card of each text block are vectorized and written into the vector database and retrieval engine to form a queryable vector index set.

[0006] Preferably, the process of constructing a hybrid index of tags and semantics is as follows: Extract entity keywords, category labels, and semantic feature vectors from the queryable vector index set; By combining entity attributes, synonyms, hierarchical relationships, and historical weak supervision signals in the business entity dictionary, a multi-level label mapping and semantic index are generated; For each entity mention, a label and semantic mapping table is created, and index priorities are generated, outputting a complete label and semantic hybrid index.

[0007] Preferably, the process of generating the candidate entity set is as follows: Retrieve candidate entity sets based on a complete hybrid index of tags and semantics, covering different text blocks and different session sources; By combining entity summary cards and context similarity, candidate entities that are duplicated, have low matching scores, or are inconsistent with the context are eliminated; Based on business rules and recall priorities, candidate entities are sorted and filtered to form an accurate and matching set of candidate entities.

[0008] Preferably, the process of generating probability scores for the first channel using a zero-shot common-index classifier is as follows: The candidate entity set and the block summary are input into a two-channel scoring model, including a first channel and a second channel; Extract the corresponding sentence and context information within the fixed windows before and after each candidate entity set from the block summary; The entity name, context text, sentence vector, keyword vector, and key tag information in the entity summary card are integrated into a multi-dimensional feature vector to form the contextual semantic representation of each entity; The contextual representation of each pair of candidate entities is mapped to a unified semantic space using a pre-trained language model; In this semantic space, the semantic similarity, contextual consistency, and entity type matching features of candidate entity pairs are calculated. A cross-block attention mechanism is introduced. For each pair of candidate entities, the zero-sample co-reference classifier outputs the reference consistency probability. Combined with entity context similarity, semantic vector inner product and type matching score, a comprehensive probability score is generated.

[0009] Preferably, the calibration process for the second channel based on business rules is as follows: The comprehensive probability score is input into the rule calibration module to extract the contextual semantic representation, entity type and historical interaction information of each pair of candidate entities; Based on a predefined set of business rules, each pair of candidate entities is verified and corrected one by one; For candidate entity pairs with low confidence or abnormal scores, a rule correction strategy is applied to adjust them, and the final output is the candidate entity pair score result after being calibrated by business rules.

[0010] Preferably, the process of weightedly fusing the two to generate the final matching score for candidate entity pairs is as follows: Input the comprehensive probability score generated by the zero-sample co-index classifier in the first channel and the candidate entity score in the second channel after calibration by business rules; The two types of scores are weighted according to preset weights; The weighted results are normalized to unify the scoring criteria; For candidate entity pairs with low confidence, secondary adjustments can be made by combining contextual similarity and entity type information to finally output a matching score.

[0011] Preferably, the process of constructing a weighted entity graph is as follows: Based on the final matching score of the candidate entity pairs, each candidate entity is treated as a node, and the block summary, entity summary card, and context vector information are recorded. The final matching score of each pair of candidate entities is used as the edge weight, and the edge weight is adjusted by combining context similarity, entity type consistency and historical weak supervision signals, so as to construct a complete weighted entity graph.

[0012] Preferably, the process of applying the maximum spanning tree algorithm to solve the global preliminary referential chain structure on a weighted entity graph is as follows: The weighted entity graph is solved by performing maximum spanning tree calculation based on edge weights to form a tree structure covering nodes, with each tree corresponding to a preliminary referential chain; During the solution process, the node order of each chain and the corresponding intra-block summary, entity summary card and context vector information are recorded; Output the global preliminary reference chain structure consisting of all preliminary reference chains.

[0013] Preferably, the process of forming a globally consistent and callable set of global referential chains is as follows: Identify nodes with multiple or undefined references in the global preliminary reference chain structure; For conflicting nodes, semantic consistency and context matching degree with candidate identical nodes are calculated by combining entity type, block summary and context vector information; Based on the calculation results, conflicting nodes are prioritized and relinked, low-confidence or semantically inconsistent connections are removed, and missing references are filled in, ultimately forming a globally consistent and callable set of global reference chains.

[0014] Compared with existing technologies, this invention provides a method for resolving cross-dialogue references based on long texts, which has the following beneficial effects: This invention intelligently segments and compresses ultra-long texts at a high ratio, preserving core entity information while significantly reducing processing redundancy. By combining in-block summaries, entity summary cards, business entity dictionaries, and historical weak supervision signals, a hybrid tag and semantic index is constructed, enabling efficient recall and filtering of candidate entities across blocks, sessions, and channels. Through a dual-channel scoring model and weighted fusion strategy, the matching scores of candidate entity pairs can be accurately calculated. Furthermore, the maximum spanning tree algorithm is applied to the weighted entity graph to form a preliminary global referential chain structure. Conflicting nodes are corrected and filled by combining entity type, in-block summary, and context vectors, thereby generating a globally consistent and callable set of global referential chains. This method effectively processes ultra-long, multi-session texts while ensuring the accuracy of referential resolution, significantly improving the accuracy and consistency of cross-dialogue entity recognition, enhancing the system's understanding of complex texts, and providing reliable basic data support for subsequent applications such as information extraction, knowledge graph construction, and intelligent question answering. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the method of the present invention. Detailed Implementation

[0016] 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.

[0017] Example 1: Please refer to Figure 1 As shown in the embodiment of the present invention, a method for resolving cross-dialogue references based on long text includes the following steps: S1: Divide the extremely long text into blocks, and segment them into text blocks according to length and semantic continuity. Entity reference sentences are retained, non-entity sentences are compressed, and block summaries and entity summary cards are generated. These are then vectorized to form a set of queryable vector indexes.

[0018] The process of generating intra-block digests and entity digest cards in S1 is as follows: Semantic analysis algorithms are used to calculate the semantic similarity between sentences in the text, and the text is segmented into blocks in combination with text length constraints; Semantic analysis is performed on the input long text, which is divided into basic units by sentence and the semantic similarity between sentences is calculated to measure the continuity of context. Combined with length constraints, such as no more than 1000 tokens per block and semantic continuity constraints, the text is divided into several blocks. While maintaining complete semantic units, each block allows a certain proportion of sliding overlap, such as 20% sentence overlap, to ensure the integrity of entity context in dialogues or long texts.

[0019] High-ratio compression is performed on non-entity sentences within the block, while preserving key contextual information during the compression process, including core vocabulary in the sentence and semantic connections between sentences; High-ratio compression (approximately 90%) is applied to non-entity sentences within each text block. During compression, core vocabulary and semantic connections between sentences are preserved. The compression strategy employs entity-aware compression, prioritizing the retention of sentences containing entities and their context to ensure the integrity of semantic information required for subsequent cross-block retrieval and referential resolution. This approach, based on semantic continuity segmentation, compression ratio, and entity summary cards, achieves efficient block segmentation and information condensation of ultra-long texts, providing an original and feasible foundation for cross-block global referential resolution.

[0020] Generate an intra-block summary for each text block, recording the original sentence position, keyword index, and sentence vector representation; Within each compressed text block, an intra-block summary is generated. This summary retains the core information within the block in the form of compressed sentences, while also including all entity references. The summary records the original position, keyword index, and sentence vector representation of each sentence, providing a vectorized semantic index for cross-block semantic retrieval and global referential resolution. This ensures that the key information of the text block is computable and accessible, while mitigating information loss caused by boundary effects.

[0021] For each entity in a text block, generate an entity summary card, including the entity name, location of occurrence, context window, and key tag information; For each entity mention within a text block, an entity summary card is generated, recording the entity name, location of occurrence, context window, and key tag information, such as entity category and business tags. The entity summary card structurally preserves the semantic and contextual features of the entity.

[0022] The intra-block summary and entity summary card of each text block are vectorized and written into the vector database and retrieval engine to form a queryable vector index set.

[0023] The intra-block summary and all entity summary cards within each text block are vectorized, including sentence vectors, keyword vectors, and context vectors. The processed vector data is then written into a vector database and retrieval engine to form a queryable vector index set. Through vectorization, efficient candidate entity retrieval across blocks, sessions, and channels is achieved, providing a foundation for ranking and global fusion.

[0024] S2: Based on a queryable set of vector indexes, combined with a business entity dictionary and historical weak supervision signals, a hybrid index of tags and semantics is constructed. Cross-block, cross-session, and cross-channel retrieval and filtering are performed on each entity mention to generate a candidate entity set.

[0025] The process of constructing a hybrid index of labels and semantics in S2 is as follows: Extract entity keywords, category labels, and semantic feature vectors from the queryable vector index set; Entity-related information is extracted from a set of queryable vector indexes that have been vectorized and stored in a vector database or retrieval engine. The features extracted for each entity include entity keywords, such as order number, mobile phone number, product name and other core business information; category tags, corresponding business entity types, such as customer information, logistics information and so on; and semantic feature vectors, which are high-dimensional vectors generated by pre-trained language models or text embedding models, reflecting the semantic representation of the entity in the context.

[0026] By combining entity attributes, synonyms, hierarchical relationships, and historical weak supervision signals in the business entity dictionary, a multi-level label mapping and semantic index are generated; By leveraging entity attributes, synonyms, hierarchical relationships, and historical weak supervision signals in the business entity dictionary, such as past referential relationships, historical annotations, or rule verification results, multi-level label mappings and semantic indexes are generated for each entity to achieve multi-dimensional recall. Level 1: Precise matching of entity name and category label; Level 2: Expansion of synonyms and hierarchical relationships to achieve semantic generalization recall; Level 3: Combining historical weak supervision signals to assign priorities or weights to potential candidates to ensure the business rationality and contextual consistency of the recall.

[0027] For each entity mention, a tag-semantic mapping table is created, and an index priority is generated, outputting a complete tag-semantic hybrid index; A tag and semantic mapping table is created for each entity mention, and an index priority is generated. The output is a complete set of mixed tag and semantic indexes. This index not only enables cross-block recall, but also enables effective matching of the same entity across sessions and channels. Through a multi-level recall strategy, precise recall, semantic recall and historical experience are combined to improve the coverage of candidate entities and solve the problem of missed candidates in traditional single-channel recall.

[0028] The process of generating the candidate entity set in S2 is as follows: Retrieve candidate entity sets based on a complete hybrid index of tags and semantics, covering different text blocks and different session sources; The system leverages a comprehensive hybrid index of tags and semantics to retrieve each entity mention, with results covering different text blocks, conversational sessions, and sources. The tag index ensures matching of entity type, business attributes, and keywords, while the semantic index uses vectorized representations to capture semantic similarity, thereby recalling potential co-referential entities in cross-block and cross-session scenarios.

[0029] By combining entity summary cards and context similarity, candidate entities that are duplicated, have low matching scores, or are inconsistent with the context are eliminated; For the initially retrieved candidate entity set, the context similarity is calculated by combining the entity occurrence position, context window and core tag information in the entity summary card. The semantic and contextual consistency between the candidate entity and the target entity is evaluated. By setting a similarity threshold, candidate entities that are duplicated, have low matching degree or inconsistent context are eliminated to ensure that the candidate set is not only comprehensive but also highly relevant to the semantics and business scenario of the target entity, thereby reducing mismatches and noise interference.

[0030] Based on business rules and recall priorities, candidate entities are sorted and filtered to form an accurate and matching set of candidate entities; After screening, the candidate entity set is sorted and scored according to business rules, such as time order, order number matching, dialogue session priority, and recall priority information. The index priority is determined by tag mapping and historical weak supervision signals to ensure that high-priority entities participate in subsequent scoring and global fusion first, resulting in an accurate and matchable candidate entity set.

[0031] S3: Input the candidate entity set and the block summary into the dual-channel scoring model. The first channel uses a zero-sample common index classifier to generate a probability score, and the second channel is calibrated based on business rules. The two are weighted and fused to generate the final matching score of the candidate entity pair.

[0032] The process of generating probability scores for the first channel in S3 using a zero-sample common-index classifier is as follows: The candidate entity set and the block summary are input into a two-channel scoring model, including a first channel and a second channel; The candidate entity set and the block summary are input into a dual-channel scoring model, which includes a first-channel zero-sample common index classifier and a second-channel business rule calibration. The candidate entity set contains high-matching entities obtained from cross-block, cross-session, and cross-channel recall stages, and the block summary includes the core contextual information of the sentence in which each entity is located.

[0033] Extract the corresponding sentence and context information within the fixed windows before and after each candidate entity set from the block summary; For each candidate entity, extract the sentence containing the entity and the context sentences in the fixed windows before and after it from the block summary. The context windows can be set according to the dialogue structure, such as 2-3 sentences before and after, to ensure that the logical relationship and semantic connection between the entity are fully captured and to solve the problem of ambiguity caused by insufficient context in a single sentence.

[0034] The entity name, context text, sentence vector, keyword vector, and key tag information in the entity summary card are integrated into a multi-dimensional feature vector to form the contextual semantic representation of each entity; The entity name, context text, sentence vector, keyword vector, and key label information from the entity summary card for each candidate entity are integrated into a multi-dimensional feature vector. The sentence vector is generated by a pre-trained language model, the keyword vector is extracted from the entity keyword index, and the label information comes from the business entity dictionary and historical weak supervision signals. Through this multi-dimensional feature integration, the model can simultaneously perceive entity semantics, contextual association, and business attributes in a unified vector space, achieve cross-block semantic alignment, and form rich semantic representations for subsequent zero-sample same-reference judgment and global reference chain construction.

[0035] The contextual representation of each pair of candidate entities is mapped to a unified semantic space using a pre-trained language model; The contextual semantic representation corresponding to each candidate entity is constructed into a standard input sequence. The input sequence includes the entity name, the context window text of the entity in the original text, the sentence vector identifier, and the key label information in the entity summary card. These are concatenated and encoded according to the input format of the pre-trained language model. The encoded input sequence is then fed into the encoding layer of the pre-trained language model. Through a multi-layer self-attention mechanism, the term relations, context dependencies, and entity semantics in the input sequence are jointly modeled, and the corresponding high-dimensional semantic embedding vectors are output. The semantic embedding vectors of different candidate entities are dimensionally aligned and normalized to map them to the same vector space, thereby ensuring the comparability of entity contextual representations from different sources, different text blocks, and different sessions. Through the above mapping process, the original heterogeneous text features are transformed into semantic feature representations of a unified scale, providing a unified semantic foundation for semantic similarity calculation, contextual consistency analysis, and referential relationship judgment between entity pairs.

[0036] In this semantic space, the semantic similarity, contextual consistency, and entity type matching features of candidate entity pairs are calculated. In a unified semantic space, semantic similarity, contextual consistency, and entity type matching characteristics are calculated for each pair of candidate entities. Semantic similarity measures the semantic closeness of candidate entities, contextual consistency reflects the logical coherence of candidate entities in the context, and entity type matching is used to constrain different categories of entities from being incorrectly matched.

[0037] A cross-block attention mechanism is introduced. For each pair of candidate entities, the zero-sample co-reference classifier outputs the reference consistency probability. The comprehensive probability score is generated by combining entity context similarity, semantic vector inner product and type matching score. A cross-block attention mechanism is introduced to perform weighted aggregation of the semantic representations of each pair of candidate entities. The zero-sample referential classifier outputs the referential consistency probability. Combined with entity context similarity, semantic vector inner product and type matching score, a comprehensive probability score is generated. The cross-block attention mechanism allows the model to aggregate relevant contextual information globally, improving the reliability and consistency of referential judgment in long texts and cross-conversation. Finally, each pair of candidate entities obtains a comprehensive probability score, providing a basis for dual-channel weighted fusion and global referential chain construction.

[0038] The calibration process for the second channel in S3 based on business rules is as follows: The comprehensive probability score is input into the rule calibration module to extract the contextual semantic representation, entity type and historical interaction information of each pair of candidate entities; The comprehensive probability score of the candidate entity pairs output from the first channel is input into the rule calibration module. The rule calibration module extracts calibration features related to rule judgment for each candidate entity pair. The calibration features include the contextual semantic representation of the candidate entity pair, entity type information, entity attribute information, entity appearance time and location, the identifier of the business object to which it belongs, and historical interaction records. Among them, the contextual semantic representation is used to reflect the semantic consistency of the entities in the original text, the entity type and attributes are used to constrain whether the candidate entities meet the business semantic consistency condition, and the historical interaction records are used to reflect the relationship between the entities in historical sessions or business processes, providing a judgmentable input basis for rule verification.

[0039] Based on a predefined set of business rules, each pair of candidate entities is verified and corrected one by one; A set of business rules is pre-configured in the rule calibration module. The set of rules includes entity attribute consistency rules, time logic constraint rules, business process consistency rules, and historical interaction consistency rules. For each pair of candidate entities, the above rules are matched and judged in sequence. For example, it is judged whether the entity types are consistent, whether the business object identifiers conflict, whether the occurrence time conforms to the business process order, and whether the historical interaction records support the same entity relationship. When a rule is not satisfied, a penalty coefficient or confidence reduction coefficient is applied to the comprehensive probability score of the candidate entity pair. In this way, while maintaining the continuity of model output, business constraints are introduced to achieve controllable correction of the model scoring results.

[0040] For candidate entity pairs with low confidence or abnormal scores, a rule correction strategy is applied to adjust them, and the final output is the candidate entity pair score result after business rule calibration. When the overall probability score of a candidate entity pair is lower than the preset confidence threshold, or when an anomaly occurs that clearly conflicts with the business rules, the rule correction strategy is triggered to perform further correction processing on the candidate entity pair, including lowering its score weight, marking it as a low-confidence candidate, or directly eliminating it. At the same time, the correction result is fused with the original model score to form the final candidate entity pair score result after being calibrated by the business rules, providing a stable and reliable input for ranking, global fusion, and reference chain construction.

[0041] The process of weighted fusion of the two in S3 to generate the final matching score of candidate entity pairs is as follows: Input the comprehensive probability score generated by the zero-sample co-index classifier in the first channel and the candidate entity score in the second channel after calibration by business rules; The system receives the comprehensive probability score output by the zero-sample same-index classifier in the first channel and the candidate entity pair score calibrated by the business rules in the second channel. The first channel score is used to characterize the same-index consistency of candidate entities in the semantic space, while the second channel score is used to reflect the rationality of candidate entities under business attributes, temporal relationships, and structural constraints. By simultaneously introducing two types of score signals from different sources and with different constraint dimensions, the fusion process includes both semantic relevance information and business consistency constraint information, thereby avoiding misjudgments caused by a single semantic model or a single rule system.

[0042] The two types of scores are weighted according to preset weights; Weight parameters are set for the first channel score and the second channel score respectively to adjust the influence ratio of semantic consistency judgment and business constraint verification in the final matching result. Linear weighted calculation is performed on the two types of scores to generate an initial value of the fusion score. This makes the semantic channel dominant when the semantic judgment credibility is high, and the business channel plays a greater role when the business rule constraint is stronger, thereby dynamically balancing the relationship between model judgment and rule constraint in different scenarios.

[0043] The weighted results are normalized to unify the scoring criteria; The initial weighted fusion score is normalized to map it to a uniform numerical range, thereby eliminating differences in numerical scales between different score sources and ensuring the comparability of scores between different candidate entity pairs. By unifying the score scale, the sorting, threshold filtering, and graph construction processes can be based on a stable and consistent score standard, avoiding sorting offsets or threshold failures caused by inconsistent score ranges.

[0044] For candidate entity pairs with low confidence, secondary adjustments can be made by combining contextual similarity and entity type information, and the final matching score is output. When the fusion score is lower than the preset reliability threshold, the contextual similarity, entity type consistency and time or business attribute constraint information between candidate entity pairs are introduced for secondary calibration and adjustment. This is used to suppress potential mismatches and enhance the ability to distinguish true referential relationships. After the secondary adjustment is completed, the obtained score is used as the final matching score output of the candidate entity pair, providing a unified and stable edge weight input for the construction of weighted entity graphs and the process of solving referential chains.

[0045] S4: Based on the final matching scores of candidate entity pairs, each candidate entity is treated as a node, and the final matching score between each pair of candidate entities is treated as the edge weight. A weighted entity graph is constructed, and the maximum spanning tree algorithm is applied to the weighted entity graph to solve the global preliminary referential chain structure.

[0046] The process of constructing a weighted entity graph in S4 is as follows: Based on the final matching score of the candidate entity pairs, each candidate entity is treated as a node, and the block summary, entity summary card, and context vector information are recorded. The candidate entity results are parsed to extract all entity instances. Each entity instance is uniquely identified as a graph node. An entity instance refers to a standardized entity representation obtained from entity recognition and normalization in the original text or data block. For each node, the corresponding data block summary information is stored to characterize the local semantic environment of the entity in the specific text block, entity summary card information is used to describe the type, attributes, source and basic semantic features of the entity, and context vector representation based on pre-trained language model encoding is used to characterize the distributed semantic features of the entity in the context. In this way, the originally discrete candidate entities are transformed into structured graph nodes, and each node is given multi-dimensional attribute information. This allows relation inference, graph propagation and clustering operations to not only rely on the edge weights between nodes, but also to use the semantic features and type constraints of the nodes themselves for auxiliary judgment.

[0047] The final matching score of each pair of candidate entities is used as the edge weight, and the edge weight is adjusted by combining context similarity, entity type consistency and historical weak supervision signals, so as to construct a complete weighted entity graph. For each candidate entity pair for which a final matching score has been calculated, an edge is established between the two corresponding entity nodes, and the final matching score is used as the initial weight of the edge to reflect the basic association strength between the two entities at the semantic and referential levels. Based on this, three types of correction factors are introduced to adjust the edge weights: first, a context similarity factor, which measures the semantic closeness of two entities in their contexts to enhance the connection weight between co-existing entities in similar contexts; second, an entity type consistency factor, which constrains unreasonable strong connections between different types of entities, such as "person" and "place," thereby reducing structural noise; and third, a historical weak supervision signal factor, which introduces empirical information about entity co-reference or association from existing labeled data or historical statistical results to empirically correct the edge weights. By weighting, fusing, and normalizing the initial edge weights with the above correction factors, the final edge weight value is obtained, and a complete weighted entity graph is constructed accordingly.

[0048] In S4, the process of applying the maximum spanning tree algorithm to solve the global preliminary referential chain structure on the weighted entity graph is as follows: The weighted entity graph is solved by performing maximum spanning tree calculation based on edge weights to form a tree structure covering nodes, with each tree corresponding to a preliminary referential chain; The constructed weighted entity graph is modeled as an undirected weighted graph, where nodes represent candidate entity instances, and edge weights represent the comprehensive score of the referential confidence between entity pairs. This score comprehensively reflects factors such as semantic similarity, contextual consistency, entity type matching, and historical weak supervision consistency, and is used to characterize the probability that two entities constitute the same referent at both the semantic and business levels. To avoid misjudgments caused by high local scores leading to overall errors in the referential structure, the referential inference problem is transformed into finding the optimal connection structure with the largest total edge weight covering all nodes across the entire graph. Thus, a maximum spanning tree is introduced as the global solution. The algorithm employs optimization techniques, introducing referential structure constraints during the solution process. These constraints include prohibiting the formation of closed-loop structures, preventing one entity from corresponding to multiple upstream referential nodes, and restricting unreasonable cross-type connections. This ensures that the generated structure satisfies the semantic consistency and structural rationality of the referential chain. Through these methods, the algorithm balances the relative confidence of each entity pair globally, avoiding the conflict propagation problem caused by pairwise greedy matching. It forms a tree structure with several covering nodes at the overall level, where each tree represents a preliminary referential chain inferred under the global optimal meaning, thereby realizing the transformation from local scoring to a globally consistent structure.

[0049] During the solution process, the node order of each chain and the corresponding intra-block summary, entity summary card and context vector information are recorded; While performing the maximum spanning tree solution, the corresponding entity node index and its order of appearance in the original text are recorded for each selected edge according to the connection order. Simultaneously, the block summary content, entity summary card information, and context semantic vector representation of the entity are stored together. This recording process not only preserves the structural relationship between entities, but also preserves the semantic and business basis for forming the relationship, which is used for subsequent conflict detection, structural interpretation, and correction. By jointly recording the node order of the referential chain and the context information, the system can trace back the formation cause of any referential chain and determine whether there are any abnormalities such as semantic jumps, type mutations, or context breaks in the chain. This provides an operable data foundation for introducing a conflict perception mechanism and local correction optimization, thereby improving the interpretability, verifiability, and stability of the overall system.

[0050] Output the global preliminary reference chain structure consisting of all preliminary reference chains; All tree-like structures output by the maximum spanning tree algorithm are organized into a global preliminary referential chain structure. Each referential chain represents a set of text entities that the system believes may point to the same real entity. This structure serves as the initial inference result under the global optimal meaning and is input to the subsequent conflict detection module to further identify semantic conflicts, type conflicts, contextual conflicts, and historical consistency conflicts. By first constructing a global preliminary optimal structure and then introducing a conflict detection and correction mechanism for secondary optimization, the system can correct local anomalies while ensuring global consistency. This forms a closed-loop optimization framework of "global optimal solution + conflict-aware correction," which avoids the short-sightedness of local greedy strategies and prevents excessive amplification of noise and error by one-time global optimization. Finally, a stable, reliable, and interpretable referential chain result is obtained.

[0051] S5: For conflicting entities appearing in the preliminary global referential chain structure, correction and filling are performed by combining entity type, block summary and context vector information to form a globally consistent and callable global referential chain set.

[0052] The process of forming a globally consistent and callable set of global reference chains in S5 is as follows: Identify nodes with multiple or undefined references in the global preliminary reference chain structure; The system performs a traversal analysis on the global preliminary referential chain structure output by the maximum spanning tree, counting the occurrence frequency of each entity node in different referential chains and its in-degree and out-degree connections. This identifies isolated nodes with one-to-many connections, multi-source pointing relationships, or no chain coverage. Nodes with multiple upstream or downstream connections are marked as nodes with multiple referential conflicts, and nodes without connections to any high-confidence entities are marked as nodes with unclear referentials. By performing structural consistency analysis on the referential structure itself, rather than relying solely on local similarity to determine conflicts, the system can discover potential semantic ambiguities and structural contradictions at the global level. This provides clear targets for targeted correction and completion, thereby preventing errors from propagating throughout the chain and improving the stability and reliability of the overall referential structure.

[0053] For conflicting nodes, semantic consistency and context matching degree with candidate identical nodes are calculated by combining entity type, block summary and context vector information; For conflicting nodes, a set of candidate referential entities is constructed, and multi-dimensional matching indicators are calculated between the node and each candidate entity. These indicators include semantic similarity based on context vectors, type consistency score based on entity summary cards, and contextual matching score based on block summary content. These multi-dimensional indicators are integrated into a unified referential consistency score, which is used to characterize the possibility that the current conflicting node and each candidate entity constitute the same real referent. By introducing a joint judgment of three types of information—semantics, structure, and business attributes—conflict resolution no longer relies solely on text similarity. Instead, it comprehensively considers the proximity of entities in the semantic space, the rationality of business attributes, and their appearance logic in specific contexts. This effectively distinguishes semantically similar but referentially different entities, reducing the risk of erroneous merging and splitting.

[0054] Based on the calculation results, conflicting nodes are prioritized and relinked, low-confidence or semantically inconsistent connections are removed, and missing references are filled in, ultimately forming a globally consistent and callable set of global reference chains. Based on the referential consistency score, candidate connections of conflicting nodes are prioritized and retained, with high consistency, high type matching, and logical consistency with the context being retained first. Connections with low confidence or semantic jumps or type conflicts are removed. For isolated nodes or broken chains formed after the removal of connections, the optimal connection is re-established based on the remaining candidate entities, thereby filling in the missing referential relationships and restoring the chain integrity. Through the closed-loop correction mechanism of "removal-reordering-completion" mentioned above, the referential chain gradually approaches the real referential structure while maintaining global consistency. Finally, a set of global referential chains that is structurally stable, semantically coherent, business-reasonable, and can be directly called by downstream systems is formed, thus providing a reliable foundation for cross-dialogue information fusion, knowledge integration, and reasoning tasks.

[0055] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for resolving referential issues across long text dialogues, characterized in that, Includes the following steps: The text is divided into blocks based on length and semantic continuity. Entity references are retained, while non-entity sentences are compressed to generate block summaries and entity summary cards. These are then vectorized to form a queryable vector index set. Based on a queryable set of vector indexes, combined with a business entity dictionary and historical weak supervision signals, a hybrid index of tags and semantics is constructed. Cross-block, cross-session, and cross-channel retrieval and filtering are performed on each entity mention to generate a candidate entity set. The candidate entity set and the block summary are input into a dual-channel scoring model. The first channel uses a zero-shot common index classifier to generate a probability score, and the second channel is calibrated based on business rules. The two are weighted and fused to generate the final matching score of the candidate entity pair. Based on the final matching scores of candidate entity pairs, each candidate entity is treated as a node, and the final matching score between each pair of candidate entities is treated as the edge weight. A weighted entity graph is constructed, and the maximum spanning tree algorithm is applied to the weighted entity graph to solve the global preliminary referential chain structure. For conflicting entities appearing in the preliminary global referential chain structure, the system combines entity type, block summary and context vector information to correct and fill in the gaps, forming a globally consistent and callable global referential chain set.

2. The method for resolving referential issues based on long-text cross-dialogue as described in claim 1, characterized in that, The process of forming a queryable set of vector indexes is as follows: Semantic analysis algorithms are used to calculate the semantic similarity between sentences in the text, and the text is segmented into blocks in combination with text length constraints; High-ratio compression is performed on non-entity sentences within the block, while preserving key contextual information during the compression process, including core vocabulary in the sentence and semantic connections between sentences; Generate an intra-block summary for each text block, recording the original sentence position, keyword index, and sentence vector representation; For each entity in a text block, generate an entity summary card, including the entity name, location of occurrence, context window, and key tag information; The intra-block summary and entity summary card of each text block are vectorized and written into the vector database and retrieval engine to form a queryable vector index set.

3. The method for resolving referential issues based on long-text cross-dialogue as described in claim 2, characterized in that, The process of building a hybrid index of tags and semantics is as follows: Extract entity keywords, category labels, and semantic feature vectors from the queryable vector index set; By combining entity attributes, synonyms, hierarchical relationships, and historical weak supervision signals in the business entity dictionary, a multi-level label mapping and semantic index are generated; For each entity mention, a label and semantic mapping table is created, and index priorities are generated, outputting a complete label and semantic hybrid index.

4. The method for resolving referential issues based on long-text cross-dialogue as described in claim 3, characterized in that, The process of generating a candidate entity set is as follows: Retrieve candidate entity sets based on a complete hybrid index of tags and semantics, covering different text blocks and different session sources; By combining entity summary cards and context similarity, candidate entities that are duplicated, have low matching scores, or are inconsistent with the context are eliminated; Based on business rules and recall priorities, candidate entities are sorted and filtered to form an accurate and matching set of candidate entities.

5. The method for resolving referential issues based on long-text cross-dialogue as described in claim 4, characterized in that, The process of generating probability scores using the zero-shot common index classifier in the first channel is as follows: The candidate entity set and the block summary are input into a two-channel scoring model, including a first channel and a second channel; Extract the corresponding sentence and context information within the fixed windows before and after each candidate entity set from the block summary; The entity name, context text, sentence vector, keyword vector, and key tag information in the entity summary card are integrated into a multi-dimensional feature vector to form the contextual semantic representation of each entity; The contextual representation of each pair of candidate entities is mapped to a unified semantic space using a pre-trained language model; In this semantic space, the semantic similarity, contextual consistency, and entity type matching features of candidate entity pairs are calculated. A cross-block attention mechanism is introduced. For each pair of candidate entities, the zero-sample co-reference classifier outputs the reference consistency probability. Combined with entity context similarity, semantic vector inner product and type matching score, a comprehensive probability score is generated.

6. The method for resolving referential issues based on long-text cross-dialogue as described in claim 5, characterized in that, The calibration process for the second channel based on business rules is as follows: The comprehensive probability score is input into the rule calibration module to extract the contextual semantic representation, entity type and historical interaction information of each pair of candidate entities; Based on a predefined set of business rules, each pair of candidate entities is verified and corrected one by one; For candidate entity pairs with low confidence or abnormal scores, a rule correction strategy is applied to adjust them, and the final output is the candidate entity pair score result after being calibrated by business rules.

7. The method for resolving referential issues based on long-text cross-dialogue as described in claim 6, characterized in that, The process of weightedly fusing the two to generate the final matching score for candidate entity pairs is as follows: Input the comprehensive probability score generated by the zero-sample co-index classifier in the first channel and the candidate entity score in the second channel after calibration by business rules; The two types of scores are weighted according to preset weights; The weighted results are normalized to unify the scoring criteria; For candidate entity pairs with low confidence, secondary adjustments can be made by combining contextual similarity and entity type information to finally output a matching score.

8. The method for resolving referential issues based on long-text cross-dialogue as described in claim 7, characterized in that, The process of constructing a weighted entity graph is as follows: Based on the final matching score of the candidate entity pairs, each candidate entity is treated as a node, and the block summary, entity summary card, and context vector information are recorded. The final matching score of each pair of candidate entities is used as the edge weight, and the edge weight is adjusted by combining context similarity, entity type consistency and historical weak supervision signals, so as to construct a complete weighted entity graph.

9. A method for resolving referential issues based on long-text cross-dialogue as described in claim 8, characterized in that, The process of applying the maximum spanning tree algorithm to solve the global preliminary referential chain structure on a weighted entity graph is as follows: The weighted entity graph is solved by performing maximum spanning tree calculation based on edge weights to form a tree structure covering nodes, with each tree corresponding to a preliminary referential chain; During the solution process, the node order of each chain and the corresponding intra-block summary, entity summary card and context vector information are recorded; Output the global preliminary reference chain structure consisting of all preliminary reference chains.

10. A method for resolving referential issues based on long-text cross-dialogue as described in claim 9, characterized in that, The process of forming a globally consistent and callable set of global reference chains is as follows: Identify nodes with multiple or undefined references in the global preliminary reference chain structure; For conflicting nodes, semantic consistency and context matching degree with candidate identical nodes are calculated by combining entity type, block summary and context vector information; Based on the calculation results, conflicting nodes are prioritized and relinked, low-confidence or semantically inconsistent connections are removed, and missing references are filled in, ultimately forming a globally consistent and callable set of global reference chains.

Images