CN-121979995-A - Cross-round semantic consistency improving method and device based on bidirectional graph index

CN121979995ACN 121979995 ACN121979995 ACN 121979995ACN-121979995-A

Abstract

The application discloses a cross-round semantic consistency improving method and device based on a bidirectional graph index, which relate to the field of artificial intelligent dialogue systems, wherein the method comprises the steps of responding to the receiving of a first dialogue fragment, entity extraction is carried out on the first dialogue fragment, the graph weight of the extracted first entity and the first dialogue fragment is determined, and then the entity-fragment index and the heterogeneous graph are constructed; the method comprises the steps of receiving dialogue fragment inquiry instructions, determining candidate entities from a first entity, determining the association degree between the candidate entities and target dialogue fragments according to graph weights and similarity matrixes of the dialogue fragments and the entities, further determining target candidate entities, and determining the comprehensive credibility of the target candidate entities according to association evidence information and version information corresponding to the target candidate entities aiming at each target candidate entity, further determining the target entities. The application can improve the consistency of semantic understanding of the user when the artificial intelligent dialogue system performs the cross-round dialogue, and improve the user experience when the artificial intelligent dialogue system performs the cross-round dialogue.

Inventors

WANG GUOZHANG
ZHONG YANG
CUI JIAN

Assignees

和元达信息科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260205

Claims (10)

1. The cross-round semantic consistency improving method based on the bidirectional graph index is characterized by comprising the following steps of: in response to receiving the first dialogue fragment, entity extraction is carried out on the first dialogue fragment, and a first entity corresponding to the first dialogue fragment and a first confidence coefficient of the first entity are determined; Determining a first graph weight of the first entity and the first dialogue fragment according to the first confidence degree and the semantic relativity of the first entity and the first dialogue fragment; constructing an entity-segment index according to the first entity, the first dialogue segment and the first graph weight, and constructing an heterogram according to the entity-segment index, wherein the entity-segment index comprises an entity and a dialogue segment corresponding to the entity, and graph weights, associated evidence information and version information corresponding to the entity and the entity; In response to receiving a dialogue fragment query instruction, determining candidate entities from the first entities according to the dialogue fragment query instruction, wherein the dialogue fragment query instruction is used for indicating to determine target entities corresponding to target dialogue fragments under a target version; Determining the association degree between the candidate entity and the target dialogue fragment according to the candidate entity, the target dialogue fragment, the first graph weight corresponding to the candidate entity, and the similarity matrix between the dialogue fragment corresponding to the heterogeneous graph and the entity; Determining a target candidate entity from the candidate entities according to the association degree between the candidate entities and the target dialogue fragments; Determining the entity-segment index corresponding to the target candidate entity according to the heterogram, determining the evidence consistency of the target candidate entity according to the associated evidence information and version information of the entity-segment index corresponding to the target candidate entity, and determining the comprehensive credibility of the target candidate entity according to the evidence consistency and association degree of the target candidate entity; And determining the target entity corresponding to the target dialogue fragment from the target candidate entities according to the comprehensive credibility of the target candidate entities.
2. The bi-graph index based cross-round semantic consistency promotion method of claim 1, the cross-round semantic consistency improving method based on the bidirectional graph index is characterized by further comprising the following steps of: In response to receiving the second dialogue fragment, performing entity extraction on the second dialogue fragment according to the second dialogue fragment, and determining a second entity corresponding to the second dialogue fragment and a second confidence of the second entity; determining a second graph weight of the second entity and the second dialogue fragment according to the second confidence degree and the semantic relativity of the second entity and the second dialogue fragment; updating an entity-segment index and an iso-composition according to the second entity, the second dialog segment and the second graph weight; And in response to the updating of the entity-segment index and the abnormal composition, updating the similarity matrix according to the second dialogue segment and the second entity through local recalculation.
3. The bi-pass semantic consistency promotion method based on bi-pass index according to claim 2, wherein before updating entity-segment index and heterogeneous map according to the second entity, the second dialog segment, the second map weight, the bi-pass semantic consistency promotion method based on bi-pass index further comprises: Determining a third entity from the second entities, wherein the third entity is the second entity belonging to the first entity; determining a change state of the third entity according to the first confidence coefficient, the second confidence coefficient and a preset change threshold value corresponding to the third entity; And determining an updating strategy of the entity-fragment index corresponding to the first entity according to the change state of the third entity.
4. The bi-graph index-based cross-round semantic consistency promotion method of claim 1, wherein the first session segment is multi-modal data, the entity extraction is performed on the first session segment, and a first entity corresponding to the first session segment and a first confidence coefficient of the first entity are determined, including: Extracting multi-mode entities from the first dialogue fragment, and determining a first entity of the first dialogue fragment in a plurality of modes and a plurality of confidence degrees of the first entity in the plurality of modes; and determining the first confidence coefficient according to the plurality of confidence coefficients of the first entity under the plurality of modes and the historical confidence coefficient of the first entity.
5. The bi-graph index-based cross-round semantic consistency promotion method of claim 4, further comprising: according to the characteristics of the first dialogue segment in a plurality of modes, determining the collision probability of the first dialogue segment in a plurality of dimensions; the collision probability of the first dialogue segment under multiple dimensions is subjected to matrix fusion to obtain a multi-mode semantic collision probability matrix; determining a conflict risk according to the multi-mode semantic conflict probability matrix and a preset conflict risk threshold; and determining a risk resolution strategy according to the conflict risk.
6. The bi-graph index-based cross-round semantic consistency promotion method of claim 5, further comprising: determining a change value of the historical conflict hit rate according to the current conflict hit rate and the historical conflict hit rate; And updating a preset conflict risk threshold according to the change value of the historical conflict hit rate.
7. The device for improving cross-round semantic consistency based on the bidirectional graph index is characterized by comprising the following components: The entity-fragment index construction unit is used for responding to the received first dialogue fragment, carrying out entity extraction on the first dialogue fragment, determining a first entity corresponding to the first dialogue fragment and a first confidence coefficient of the first entity, determining a first graph weight of the first entity and the first dialogue fragment according to the first confidence coefficient and the semantic relativity of the first entity and the first dialogue fragment, constructing an entity-fragment index according to the first entity, the first dialogue fragment and the first graph weight, constructing an abnormal graph according to the entity-fragment index, and constructing an abnormal graph according to the entity-fragment index, wherein the entity-fragment index comprises the dialogue fragment corresponding to the entity and the entity, and graph weights corresponding to the entity, associated evidence information and version information; The segment-entity index generation unit is used for responding to a received dialogue segment query instruction, determining candidate entities from the first entity according to the dialogue segment query instruction, determining target entities corresponding to target dialogue segments under a target version according to the dialogue segment query instruction, determining the association degree between the candidate entities and the target dialogue segments according to the candidate entities, the target dialogue segments, the first graph weight corresponding to the candidate entities and the similarity matrix between dialogue segments corresponding to the heterogeneous graph and the entities, determining target candidate entities from the candidate entities according to the association degree between the candidate entities and the target dialogue segments, determining entity-segment indexes corresponding to the target candidate entities according to the heterograms for each target candidate entity, determining the evidence consistency of the target candidate entities according to the associated evidence information and version information of the entity-segment indexes corresponding to the target candidate entities, determining the comprehensive credibility of the target candidate entities according to the evidence consistency and the association degree of the target candidate entities, and determining the target entities corresponding to the target dialogue segments from the target candidate entities according to the comprehensive credibility of the target candidate entities.
8. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the bipartite graph index based cross-round semantic consistency promotion method according to any of claims 1-6.
9. A computer readable storage medium having stored thereon a computer program, which when executed by a processor implements the steps of the bi-graph index based cross-round semantic consistency promotion method of any of claims 1-6.
10. A computer program product comprising a computer program, characterized in that the computer program when executed by a processor implements the steps of the cross-round semantic consistency improving method based on bipartite graph indexing according to any of claims 1-6.

Description

Cross-round semantic consistency improving method and device based on bidirectional graph index Technical Field The application relates to the field of artificial intelligence dialogue systems, in particular to a cross-round semantic consistency improving method and device based on a bidirectional graph index. Background Multi-round dialog systems generally face challenges that severely limit the performance and user experience of the system in practical applications. For example, the state fragmentation problem manifests itself as difficulty in maintaining the uniformity and consistency of user intent in successive rounds of interaction with conventional sequence-based models, resulting in breaks in dialog states between different rounds. Secondly, semantic drift phenomenon is prominent, especially when a large language model is used to process a long dialogue, the model is easy to generate replies which are logically contradictory or deviate from the original semantics, thereby affecting the accuracy and reliability of the dialogue. For these challenges, existing solutions, such as a memory network, a unidirectional index model, and a static semantic alignment algorithm, often have difficulty in supporting dynamic backtracking capability effectively, and cannot fundamentally solve the context continuity problem in long dialogs, resulting in poor user experience when dialogs are performed across rounds. Disclosure of Invention The application aims to provide a cross-round semantic consistency improving method and device based on a bipartite graph index, which can improve consistency of semantic understanding of an artificial intelligent dialogue system and a user during cross-round dialogue and improve user experience during cross-round dialogue. In order to achieve the above object, the present application provides the following solutions: In a first aspect, the present application provides a cross-round semantic consistency improving method based on bidirectional graph indexes, including: in response to receiving a first dialog segment, entity extraction is performed on the first dialog segment, a first entity corresponding to the first dialog segment and a first confidence level of the first entity are determined, a first graph weight of the first entity and the first dialog segment is determined according to the first confidence level and the semantic relativity of the first entity and the first dialog segment, an entity-segment index is constructed according to the first entity, the first dialog segment and the first graph weight, an abnormal graph is constructed according to the entity-segment index, the entity-segment index comprises entity and entity corresponding dialog segments, and graph weights and associated evidence information and version information corresponding to the entity and the entity, a candidate entity is determined from the first entity according to a dialog segment query instruction, the dialog segment query instruction is used for indicating and determining a target entity corresponding to the target dialog segment under the target according to the first confidence level and the target dialog segment, the degree of relativity between the candidate entity and the target dialog segment is determined according to the similarity matrix between the candidate entity, the candidate entity and the target dialog segment, the entity is determined from the entity corresponding to the entity, the candidate entity can be determined according to the entity corresponding to the entity, the candidate entity can be correlated with the candidate entity, the candidate entity can be determined from the candidate entity according to the entity corresponding evidence, the candidate entity can be correlated with the candidate entity, the candidate entity can be determined from the candidate entity, the candidate has the candidate entity corresponding evidence, and the candidate has the target version according to the candidate evidence, and the candidate entity, and the candidate has the target state, and determining a target entity corresponding to the target dialogue fragment. In a second aspect, the present application provides a device for improving cross-round semantic consistency based on bidirectional graph indexing, including: The entity-fragment index construction unit is used for responding to the received first dialogue fragment, carrying out entity extraction on the first dialogue fragment, determining a first entity corresponding to the first dialogue fragment and a first confidence coefficient of the first entity, determining a first graph weight of the first entity and the first dialogue fragment according to the first confidence coefficient and the semantic relativity of the first entity and the first dialogue fragment, constructing an entity-fragment index according to the first entity, the first dialogue fragment and the first graph weight, and constructing an abnormal graph accordin