CN-121998106-A - Intelligent interaction-oriented memory enhancement and reasoning method and system

Abstract

The application provides a memory enhancement and reasoning method and system oriented to intelligent interaction, which belong to the technical field of intelligent interaction, and are characterized in that an original interaction data stream of a user is captured, segmented and arranged into an interaction event time sequence, and interaction intention is primarily judged and integrated into an interaction round unit sequence. And then extracting core keywords, and obtaining a background support knowledge unit by association expansion to generate an interaction situation knowledge set. And then, the context knowledge coding vector is semantically fused with the interactive content, and the history memory entry is searched in a long-term interactive memory bank. And finally, completing the current interaction intention according to the history memory, generating response reasoning content and transmitting the response reasoning content to the intelligent interaction terminal for presentation. The application improves the accuracy and the user experience of intelligent interaction.

Inventors

• TONG LIANG
• WANG KAIXUAN
• Mai Shaoxiong

Assignees

• 上海明奇网络科技有限公司

Dates

Publication Date

20260508

Application Date

20260410

Claims (10)

1. 1. The intelligent interaction-oriented memory enhancement and reasoning method is characterized by comprising the following steps of: Capturing a user original interaction data stream generated by an intelligent interaction terminal in a continuous interaction process, performing interaction event segmentation processing on the user original interaction data stream to obtain a plurality of interaction event units, endowing each interaction event unit with an event occurrence time tag, and performing timing sequence arrangement processing on the interaction event units according to the event occurrence time tags to generate an interaction event timing sequence; analyzing the interaction content semantics and the interaction operation type of each interaction event unit in the interaction event time sequence, performing interaction intention preliminary judgment processing on the interaction event units according to the interaction content semantics and the interaction operation type to obtain preliminary intention labels, and performing interaction turn merging processing on the interaction event units according to the event occurrence time labels and the preliminary intention labels to generate an interaction turn unit sequence carrying intention identifications; extracting a core interaction appeal keyword and a core interaction object keyword of each interaction round unit in the interaction round unit sequence, carrying out knowledge association expansion processing in a preset interaction knowledge base according to the core interaction appeal keyword to obtain a background support knowledge unit, and carrying out object anchoring processing on the background support knowledge unit according to the core interaction object keyword to generate an interaction situation knowledge set corresponding to the interaction round unit; Inputting the interaction situation knowledge set into a memory-enhanced reasoning network to perform situation knowledge coding processing to generate a situation knowledge coding vector, performing semantic fusion processing on the situation knowledge coding vector and interaction content semantics of the interaction round unit to generate a fusion memory characterization vector, and retrieving and acquiring a history interaction round memory item matched with the current interaction round in a long-term interaction memory base according to the fusion memory characterization vector; Extracting historical interaction appeal keywords and historical interaction object keywords from the historical interaction round memory items, carrying out intention completion processing on the core interaction appeal keywords and the core interaction object keywords of the current interaction round according to the historical interaction appeal keywords and the historical interaction object keywords to generate enhanced interaction intention characterization, generating response reasoning content aiming at the current interaction round according to the enhanced interaction intention characterization, and transmitting the response reasoning content to the intelligent interaction terminal to execute interaction content presentation processing.
2. 2. The intelligent interaction-oriented memory enhancement and reasoning method according to claim 1, wherein the analyzing the interaction content semantics and the interaction operation type of each interaction event unit in the interaction event time sequence, performing the preliminary determination processing of the interaction intention on the interaction event unit according to the interaction content semantics and the interaction operation type to obtain a preliminary intention tag, performing the interactive turn merging processing on the interaction event unit according to the event occurrence time tag and the preliminary intention tag to generate an interaction turn unit sequence carrying an intention identifier, includes: Carrying out multi-mode content analysis processing on each interaction event unit, extracting text semantic content, touch semantic content and sight semantic content contained in the interaction event unit, and carrying out semantic alignment fusion processing on the text semantic content, the touch semantic content and the sight semantic content to construct multi-mode interaction content semantic representation corresponding to the interaction event unit; Performing semantic classification processing on the interactive event unit according to the multi-mode interactive content semantic representation, determining a semantic category label to which the interactive event unit belongs, performing operation type identification processing on a touch operation event contained in the interactive event unit, and determining the interactive operation type of the interactive event unit according to an operation gesture type; Inputting the semantic category label and the interactive operation type into a pre-constructed intention preliminary judgment model for comprehensive reasoning treatment, wherein the intention preliminary judgment model carries out rule matching treatment on the semantic category label and the interactive operation type according to a preset intention judgment rule base, and when the semantic category label and the interactive operation type simultaneously match a certain intention judgment rule, the intention label corresponding to the intention judgment rule is used as a preliminary intention label of the interactive event unit; After each interaction event unit in the interaction event time sequence is endowed with the preliminary intention label, carrying out primary clustering treatment on the interaction event unit according to the preliminary intention label, clustering a plurality of continuous interaction event units with the same preliminary intention label into candidate interaction round groups, and distributing corresponding candidate round identifiers for each candidate interaction round group; Performing round continuity detection processing on the time interval between adjacent interaction event units in each candidate interaction round group, and dividing the adjacent interaction event units exceeding the round interruption time threshold into different candidate interaction round groups when the time interval between the adjacent interaction event units in the candidate interaction round group exceeds the preset round interruption time threshold; Reassigning candidate round identifiers for each candidate round group after segmentation, carrying out round merging and recombination processing on all the interactive event units in the interactive event time sequence according to the reassigned candidate round identifiers, and combining the interactive event units belonging to the same candidate round identifier into the same interactive round unit; Generating a corresponding turn starting time label and turn ending time label for each interactive turn unit, taking the intention label with highest frequency of occurrence in the preliminary intention labels of all interactive event units in the interactive turn unit as the intention label of the interactive turn unit, and generating the interactive turn unit sequence finally comprising the turn starting time label, the turn ending time label and the intention label.
3. 3. The intelligent interaction-oriented memory enhancement and reasoning method according to claim 1, wherein the extracting the core interaction appeal keywords and the core interaction object keywords of each interaction round unit in the interaction round unit sequence, performing knowledge association expansion processing in a preset interaction knowledge base according to the core interaction appeal keywords to obtain a background support knowledge unit, performing object anchoring processing on the background support knowledge unit according to the core interaction object keywords to generate an interaction context knowledge set corresponding to the interaction round unit, includes: Carrying out multi-mode content integration processing on each interactive round unit, and carrying out aggregation and integration on text semantic content, touch semantic content and sight semantic content of all interactive event units contained in the interactive round unit to generate complete interactive content text of the interactive round unit; Dividing the complete interactive content text into a plurality of text sentence units by sentence dividing processing, dividing each text sentence unit into a plurality of text vocabulary units by word dividing processing, determining the part-of-speech category of each text vocabulary unit by part-of-speech tagging processing, and identifying entity mention information in each text vocabulary unit by named entity identification processing; Screening noun vocabulary units and verb vocabulary units from all text vocabulary units according to the part-of-speech tagging processing result to serve as candidate key vocabulary units, and carrying out word frequency statistics processing on the candidate key vocabulary units to calculate the occurrence frequency of each candidate key vocabulary unit in the complete interactive content text; Performing descending order arrangement processing on the candidate key word units according to the occurrence frequency, selecting a preset number of candidate key word units with the top ranking of the occurrence frequency as preliminary screening key words, performing semantic role marking processing on the preliminary screening key words, and determining semantic role types of each preliminary screening key word in the interactive round unit; Identifying vocabulary units belonging to a appeal action role from the preliminary screening keyword according to the semantic role type as the core interaction appeal keywords, and identifying vocabulary units belonging to a appeal object role from the preliminary screening keyword according to the semantic role type as candidate core interaction object keywords; Performing entity link processing on the candidate core interaction object keywords, mapping each candidate core interaction object keyword to a corresponding object node in a preset interaction object knowledge base, and acquiring object attribute description information corresponding to each candidate core interaction object keyword; performing importance evaluation processing on the candidate core interaction object keywords according to the object attribute description information, calculating the key degree score of each candidate core interaction object keyword in the interaction round unit, and selecting the candidate core interaction object keywords with the highest key degree score as the core interaction object keywords; Inputting the core interaction appeal keywords into the preset interaction knowledge base for concept matching processing, searching appeal concept nodes with semanteme similar to the core interaction appeal keywords in the interaction knowledge base, and collecting a plurality of related knowledge nodes in the interaction knowledge base along a preset knowledge association relation path from the appeal concept nodes; Performing de-duplication filtering processing on all collected related knowledge nodes, performing relevance scoring processing on the de-duplicated related knowledge nodes to calculate semantic relevance scores between each related knowledge node and the appeal concept nodes, and selecting a preset number of related knowledge nodes with the semantic relevance scores ranked at the front as candidate background supporting knowledge units; Inputting the key words of the core interaction objects into the interaction knowledge base for entity matching processing, searching object entity nodes with similar meanings to the key words of the core interaction objects in the interaction knowledge base, and obtaining object attribute information which is stored in the interaction knowledge base in an associated mode by the object entity nodes; performing object correlation filtering processing on the candidate background support knowledge units according to the object attribute information, calculating the semantic association degree between each candidate background support knowledge unit and the object entity node, and reserving the candidate background support knowledge units with the semantic association degree exceeding a preset threshold as background support knowledge units associated with the core interaction object keywords; And carrying out structural organization treatment on all reserved background support knowledge units, classifying and grouping each background support knowledge unit according to the association relation type of each background support knowledge unit and the appeal concept node, adding a corresponding object anchoring mark for each background support knowledge unit, and combining all background support knowledge units after classifying and grouping and adding the object anchoring mark into the interaction situation knowledge set.
4. 4. The intelligent interaction-oriented memory enhancement and reasoning method according to claim 1, wherein the step of inputting the interaction scenario knowledge set into a memory enhancement reasoning network to perform scenario knowledge coding processing to generate scenario knowledge coding vectors, performing semantic fusion processing on the scenario knowledge coding vectors and interaction content semantics of the interaction round units to generate fusion memory characterization vectors, and retrieving and obtaining historical interaction round memory entries matched with current interaction rounds in a long-term interaction memory base according to the fusion memory characterization vectors comprises the steps of: Performing text content extraction processing on each background support knowledge unit in the interaction situation knowledge set, obtaining a knowledge description text corresponding to each background support knowledge unit, and performing text cleaning processing and sentence segmentation processing on the knowledge description text to divide the knowledge description text into a plurality of knowledge description sentence units; Inputting each knowledge description sentence unit into a knowledge encoder of the memory enhancement reasoning network to perform semantic vectorization processing, and performing vocabulary embedding mapping processing and self-attention mechanism encoding processing on the knowledge description sentence units by the knowledge encoder to generate sentence semantic vectors corresponding to each knowledge description sentence unit; carrying out aggregation processing on statement semantic vectors of all knowledge description statement units corresponding to the same background support knowledge unit, calculating average value vectors of all statement semantic vectors as knowledge unit vectors of the background support knowledge unit, and carrying out splicing and combining processing on the knowledge unit vectors of all background support knowledge units to generate context knowledge coding vectors corresponding to the interaction context knowledge set; Extracting the interaction content semantics of the interaction round units corresponding to the current interaction round in the interaction round unit sequence, storing the interaction content semantics in a multi-mode interaction content semantic representation mode, inputting the multi-mode interaction content semantic representation into a content encoder of the memory enhancement reasoning network for semantic coding processing, and converting the multi-mode interaction content semantic representation into interaction content semantic vectors with the same dimensionality as the context knowledge coding vectors; Inputting the context knowledge coding vector and the interaction content semantic vector into a cross-modal fusion module of the memory enhancement reasoning network to perform feature interaction processing, calculating attention weight distribution between the context knowledge coding vector and the interaction content semantic vector, and performing weighted adjustment processing on the interaction content semantic vector according to the attention weight distribution to generate an attention weighted interaction content semantic vector; Vector splicing is carried out on the context knowledge coding vector and the interaction content semantic vector weighted by the attention to generate a spliced fusion vector, linear transformation processing and activation function processing are carried out on the spliced fusion vector, and the dimension of the spliced fusion vector is mapped to a preset memory characterization dimension space to generate the fused memory characterization vector; Performing vectorization pre-indexing processing on each history interaction round memory item stored in the long-term interaction memory bank, inputting the history interaction round memory item into the memory enhancement reasoning network for coding processing when each history interaction round memory item is stored in the long-term interaction memory bank, and generating a history memory characterization vector corresponding to the history interaction round memory item; The history memory characterization vector and the storage address of the history interaction round memory entry are associated and stored in a vector index area of the long-term interaction memory bank, and an approximate nearest neighbor search index based on space division is constructed according to all the history memory characterization vectors stored in the vector index area; Inputting the fusion memory characterization vector of the current interaction round into the approximate nearest neighbor search index for quick search processing, determining a target subspace to which the fusion memory characterization vector belongs, and traversing and calculating vector distance measurement between the fusion memory characterization vector and each history memory characterization vector in a vector inverted index table of the target subspace; selecting a first preset number of history memory characterization vectors with the minimum vector distance measurement as preliminary candidate memory vectors, and reading complete content data of history interaction round memory entries corresponding to each preliminary candidate memory vector from the long-term interaction memory library according to the memory addresses associated and stored by the preliminary candidate memory vectors; and performing relevance reordering processing on each read history interaction round memory item, calculating a matching degree score between each history interaction round memory item and the current interaction round, and selecting a second preset number of history interaction round memory items with the highest matching degree score as history interaction round memory items matched with the current interaction round.
5. 5. The intelligent interaction-oriented memory enhancement and reasoning method according to claim 1, wherein extracting a history interaction appeal keyword and a history interaction object keyword from the history interaction round memory entry, performing intention completion processing on the core interaction appeal keyword and the core interaction object keyword of the current interaction round according to the history interaction appeal keyword and the history interaction object keyword to generate enhanced interaction intention characterization, generating response reasoning content for the current interaction round according to the enhanced interaction intention characterization, and transmitting the response reasoning content to the intelligent interaction terminal to perform interaction content presentation processing, wherein the method comprises the steps of: Analyzing the complete content data of the history interaction round memory entry, directly extracting the stored history interaction appeal keyword field from the history interaction round memory entry, and obtaining the core appeal keyword information expressed by the user in the history interaction process; Directly extracting the stored historical interaction object keyword field from the historical interaction round memory entry, acquiring core object keyword information focused by a user in the historical interaction process, performing semantic generalization processing on the historical interaction appeal keyword, and mapping the historical interaction appeal keyword to a corresponding appeal type node in a preset appeal type classification system; Acquiring the upper-level appeal category information and the lower-level appeal subclass information of the corresponding appeal type node, constructing a appeal type hierarchical tree structure of the history interaction appeal keywords, performing attribute expansion processing on the history interaction object keywords, and inquiring a complete object attribute set corresponding to the history interaction object keywords in a preset interaction object knowledge base; Performing a appeal continuation detection process on the core interaction appeal keywords and the history interaction appeal keywords of the current interaction round, calculating semantic similarity between the core interaction appeal keywords and the history interaction appeal keywords, and judging that the current interaction appeal and the history interaction appeal have a appeal continuation relation when the semantic similarity exceeds a preset appeal continuation judgment threshold; When judging that the current interaction appeal and the historical interaction appeal have the appeal continuation relation, taking the historical interaction appeal keywords as main appeal keywords, taking the core interaction appeal keywords of the current interaction rounds as sub-appeal keywords, and determining the appeal level inclusion relation between the main appeal keywords and the sub-appeal keywords according to the appeal type level tree structure of the historical interaction appeal keywords; Performing object association detection processing on the core interaction object keywords and the history interaction object keywords of the current interaction round, and identifying whether a direct or indirect object association relationship exists between the core interaction object keywords and the history interaction object keywords of the current interaction round, wherein the object association relationship comprises the same object relationship, the same category relationship and the function complementary relationship; Performing intention complement processing on the core interaction appeal keywords and the core interaction object keywords of the current interaction round according to the object relevance detection processing result and the appeal continuity detection processing result, and extracting missing appeal semantic information from the history interaction appeal keywords to complement when the core interaction appeal keywords have semantic missing, so as to obtain the complemented core interaction appeal keywords; when the attribute of the core interaction object keyword is missing, extracting missing object attribute information from a complete object attribute set of the history interaction object keyword to supplement, and carrying out combined coding processing on the supplemented core interaction appeal keyword and the core interaction object keyword to generate an enhanced interaction intention representation fused with the history interaction information; The enhanced interactive intention characterization is input into a pre-constructed response content generation model to carry out reasoning generation processing, a matched response template is searched in a preset response template library according to the completed core interactive appeal keyword, and a response template with highest appeal type matching degree with the completed core interactive appeal keyword is selected as a candidate response template; Performing variable filling processing on the fillable variable part in the candidate response template according to the completed core interaction object keywords and the description information of the object association relation, and filling the completed core interaction object keywords and the associated object information thereof into corresponding fillable variable positions to generate preliminary response text contents; Performing fluency optimization processing on the preliminary response text content, performing resolution processing on the reference relation in the preliminary response text content according to the description information of the appeal continuation relation, and performing adjustment processing on the sentence pattern structure of the preliminary response text content according to a preset language expression rule to generate response reasoning content conforming to natural language expression habit; The response reasoning content is transmitted to a display module of the intelligent interactive terminal to execute text presentation processing, the response reasoning content is displayed in a dialogue bubble mode on an interactive interface of the intelligent interactive terminal, and meanwhile, a voice synthesis module which transmits the response reasoning content to the intelligent interactive terminal executes voice synthesis processing to generate a corresponding response voice signal and plays the corresponding response voice signal through a loudspeaker; And carrying out associated storage processing on the response reasoning content and the enhanced interaction intention representation of the current interaction turn, combining and packaging the response reasoning content, the enhanced interaction intention representation and the interaction event time sequence of the current interaction turn into a new historical interaction turn memory item, and storing the new historical interaction turn memory item into the long-term interaction memory.
6. 6. The intelligent interaction-oriented memory enhancement and reasoning method according to claim 4, further comprising the step of performing memory attenuation and reinforcement processing on the long-term interaction memory bank, specifically comprising: Configuring an initial memory strength parameter and a latest access time label for each history interaction round memory item in the long-term interaction memory library, and updating the latest access time label of the history interaction round memory item as the current time when the history interaction round memory item is retrieved and output as the history interaction round memory item matched with the current interaction round; Performing incremental strengthening operation on the memory intensity parameter of the history interaction round memory entries, increasing a preset intensity increment value, and calculating the time interval length of the history interaction round memory entries from the current time according to the latest access time label of each history interaction round memory entry; Performing periodic attenuation operation on the memory intensity parameter of each history interaction round memory entry according to the time interval length, calculating an attenuation coefficient according to a preset attenuation function, and multiplying the memory intensity parameter by the attenuation coefficient to obtain an attenuated memory intensity parameter; Normalizing the memory intensity parameters of all the history interaction round memory entries in the long-term interaction memory library, and mapping the memory intensity parameters of each history interaction round memory entry into a uniform intensity value interval to obtain normalized memory intensity parameters; Screening the historical interaction round memory entries stored in the long-term interaction memory base according to the normalized memory intensity parameter, and removing the historical interaction round memory entries with the normalized memory intensity parameter lower than a preset memory elimination threshold value from the long-term interaction memory base; Deleting the history memory characterization vector corresponding to the removed history interaction round memory entries from the vector index area, performing memory integration processing on the history interaction round memory entries stored in the long-term interaction memory bank, and identifying a plurality of history interaction round memory entries with similar history interaction appeal keywords and similar history interaction object keywords; Combining the identified multiple similar history interaction round memory items to generate a new integrated memory item, wherein the history interaction complaint keywords of the integrated memory item are the union of the history interaction complaint keywords of the multiple similar history interaction round memory items; The history interaction object keywords of the integrated memory items are the union of the history interaction object keywords of a plurality of similar history interaction round memory items, and the memory strength parameters of the integrated memory items are the maximum value of the memory strength parameters of the similar history interaction round memory items; Storing the new integrated memory entry into the long-term interaction memory bank, and generating a corresponding history memory characterization vector for the new integrated memory entry to update a vector index area of the long-term interaction memory bank.
7. 7. The intelligent interaction-oriented memory enhancement and reasoning method of claim 1, further comprising the step of online updating and optimizing the memory enhancement reasoning network, specifically comprising: Collecting the enhanced interaction intention characterization and the response reasoning content corresponding to the current interaction round as a current training sample, taking the enhanced interaction intention characterization as an input characteristic, taking the response reasoning content as an expected output label, and storing the current training sample into an online training sample buffer pool; When the number of training samples stored in the online training sample buffer pool reaches a preset batch training threshold, randomly extracting training samples with preset batch size from the online training sample buffer pool to form a batch training data set; Inputting the enhanced interactive intention characterization of each training sample in the batch training data set into the memory enhanced reasoning network for forward propagation processing to obtain corresponding predicted response reasoning contents, and calculating a loss function value between the predicted response reasoning contents of each training sample and the expected output label; carrying out average processing on the loss function values of all training samples in the batch training data set according to the loss function values to obtain average loss values, and calculating gradient information of each network layer in the memory enhancement reasoning network through a back propagation algorithm according to the average loss values; updating and optimizing the network weight parameters of the memory enhancement reasoning network according to the gradient information and the preset learning rate parameters, and gradually adjusting the network weight parameters by adopting a random gradient descent optimizing algorithm to reduce the average loss value; After the network weight parameter updating and optimizing process of the memory enhancement reasoning network is completed each time, removing a batch training data set used in the updating and optimizing process from the online training sample buffer pool, and continuously storing a newly acquired current training sample into the online training sample buffer pool for subsequent batch training; Performing performance evaluation processing on an encoder part and a cross-modal fusion module of the memory enhancement reasoning network at regular intervals, inputting a preset evaluation test sample set into the memory enhancement reasoning network of the current version to perform reasoning processing, and obtaining corresponding evaluation response reasoning contents; And comparing the evaluation response reasoning content with the standard response reasoning content in the evaluation sample set, calculating the evaluation accuracy, triggering a network retraining process when the evaluation accuracy is lower than a preset accuracy threshold, and performing global retraining optimization on the memory enhancement reasoning network.
8. 8. The intelligent interaction-oriented memory enhancement and reasoning method according to claim 1, further comprising the step of dynamically expanding and updating the interaction knowledge base, specifically comprising: The method comprises the steps of monitoring a user original interaction data stream received by the intelligent interaction terminal in real time, and identifying novel interaction appeal expressions and novel interaction object references which are not covered by a current interaction knowledge base from the user original interaction data stream; Carrying out semantic clustering processing on the novel interaction demand expressions, clustering a plurality of novel interaction demand expressions expressing the same or similar demands into the same novel demand category, and generating a corresponding demand category identifier for each novel demand category; Performing entity recognition processing on the novel interaction object references, recognizing the corresponding object names, object types and object attribute information of the novel interaction object references, and generating corresponding novel object entity identifiers for each novel interaction object reference; Transmitting the novel appeal category and the novel object entity identifier to a manual auditing terminal for auditing and confirming, and receiving an auditing and confirming result returned by the manual auditing terminal, wherein the auditing and confirming result comprises a appeal category validity mark and an object entity validity mark; When the demand category validity mark is valid, the novel demand category is used as a new demand concept node to be added into the interaction knowledge base, and the demand definition description information and the demand category attribute information are configured for the new demand concept node; When the object entity validity mark is valid, adding the object entity corresponding to the novel object entity identifier as a new object entity node into the interaction knowledge base, and configuring object attribute information for the new object entity node; Establishing a appeal-object association relationship between the new appeal concept node and the new object entity node according to the co-occurrence relationship between the novel interaction appeal expression and the novel interaction object mention in the user original interaction data stream; and carrying out association relation mining processing on the existing appeal concept nodes and the existing object entity nodes in the interaction knowledge base at regular intervals, identifying potential new association relations between the existing appeal concept nodes and the existing object entity nodes, and adding the identified new association relations into the interaction knowledge base.
9. 9. The intelligent interaction-oriented memory augmentation and reasoning method of claim 1, further comprising the step of performing memory conflict detection and resolution on the long-term interaction memory bank, comprising: Before a new history interaction round memory item is stored in the long-term interaction memory library, performing similarity calculation processing on the enhanced interaction intention characterization of the new history interaction round memory item and a history memory characterization vector of the existing history interaction round memory item in the long-term interaction memory library; When the similarity of the enhanced interactive intention characterization of the new historical interaction round memory entry exceeds a preset conflict detection threshold value, marking the existing historical interaction round memory entry as a conflict candidate memory entry; Extracting the core interaction appeal keywords and the core interaction object keywords of the new history interaction round memory items, extracting the history interaction appeal keywords and the history interaction object keywords of the conflict candidate memory items, and carrying out appeal consistency comparison processing on the core interaction appeal keywords and the history interaction appeal keywords of the conflict candidate memory items; Performing object consistency comparison processing on the core interaction object keywords and the history interaction object keywords of the two, and judging that the new history interaction round memory entry and the conflict candidate memory entry are repeated memory entries when the demand consistency comparison processing result and the object consistency comparison processing result are consistent; When the repeated memory items are judged, comparing the memory intensity parameters of the new historical interaction round memory items with the memory intensity parameters of the conflict candidate memory items, reserving the historical interaction round memory items with larger memory intensity parameters, and removing the historical interaction round memory items with smaller memory intensity parameters from the long-term interaction memory library; When the resort consistency comparison processing result and the object consistency comparison processing result indicate inconsistency, judging that the new historical interaction round memory entry and the conflict candidate memory entry have memory conflict, and marking the new historical interaction round memory entry and the conflict candidate memory entry as to-be-resolved conflict memory entries; The to-be-resolved conflict memory item and the corresponding enhanced interactive intention representation and history memory representation vector are sent to a manual resolution terminal to be subjected to conflict resolution processing, and a conflict resolution result returned by the manual resolution terminal is received, wherein the conflict resolution result comprises a reserved item identifier and a removed item identifier; And carrying out reservation and removal operation on the memory items to be resolved in conflict according to the reservation item identification and the removal item identification, and updating the memory item storage state in the long-term interaction memory bank.
10. 10. An intelligent interaction-oriented memory augmentation and reasoning system comprising a processor and a computer-readable storage medium storing machine-executable instructions that, when executed by the processor, implement the intelligent interaction-oriented memory augmentation and reasoning method of any of claims 1-9.

Description

Intelligent interaction-oriented memory enhancement and reasoning method and system Technical Field The application relates to the technical field of intelligent interaction, in particular to a memory enhancement and reasoning method and system for intelligent interaction. Background In the field of intelligent interaction, with increasing complexity of interaction scenarios and increasing interaction frequency, how to accurately understand user intent and provide appropriate responses becomes a key challenge. When processing user interaction data, the existing intelligent interaction method only focuses on the information of the current interaction round, and effective utilization of historical interaction information is lacking. This results in difficulty for the system to accurately grasp the user's real intent in the face of ambiguous user expressions, ambiguous or complex interactive scenarios. For example, in a multi-turn dialog scenario, the user may mention the relevant topics in different turns, but with different emphasis and detail for each presentation. The conventional method cannot integrate the information scattered in different rounds, so that key clues are easily ignored, and the complete intention of a user cannot be accurately understood. In addition, for some interactions with implicit intentions, the existing methods lack an effective reasoning mechanism, so that potential demands of users are difficult to mine from limited interaction information, the accuracy and fluency of the interactions are seriously affected, and the increasing demands of the users on the intelligent interaction system cannot be met. Disclosure of Invention Therefore, the application aims to provide a memory enhancement and reasoning method and system oriented to intelligent interaction. According to a first aspect of the present application, there is provided a memory enhancement and reasoning method for intelligent interaction, the method comprising: Capturing a user original interaction data stream generated by an intelligent interaction terminal in a continuous interaction process, performing interaction event segmentation processing on the user original interaction data stream to obtain a plurality of interaction event units, endowing each interaction event unit with an event occurrence time tag, and performing timing sequence arrangement processing on the interaction event units according to the event occurrence time tags to generate an interaction event timing sequence; analyzing the interaction content semantics and the interaction operation type of each interaction event unit in the interaction event time sequence, performing interaction intention preliminary judgment processing on the interaction event units according to the interaction content semantics and the interaction operation type to obtain preliminary intention labels, and performing interaction turn merging processing on the interaction event units according to the event occurrence time labels and the preliminary intention labels to generate an interaction turn unit sequence carrying intention identifications; extracting a core interaction appeal keyword and a core interaction object keyword of each interaction round unit in the interaction round unit sequence, carrying out knowledge association expansion processing in a preset interaction knowledge base according to the core interaction appeal keyword to obtain a background support knowledge unit, and carrying out object anchoring processing on the background support knowledge unit according to the core interaction object keyword to generate an interaction situation knowledge set corresponding to the interaction round unit; Inputting the interaction situation knowledge set into a memory-enhanced reasoning network to perform situation knowledge coding processing to generate a situation knowledge coding vector, performing semantic fusion processing on the situation knowledge coding vector and interaction content semantics of the interaction round unit to generate a fusion memory characterization vector, and retrieving and acquiring a history interaction round memory item matched with the current interaction round in a long-term interaction memory base according to the fusion memory characterization vector; Extracting historical interaction appeal keywords and historical interaction object keywords from the historical interaction round memory items, carrying out intention completion processing on the core interaction appeal keywords and the core interaction object keywords of the current interaction round according to the historical interaction appeal keywords and the historical interaction object keywords to generate enhanced interaction intention characterization, generating response reasoning content aiming at the current interaction round according to the enhanced interaction intention characterization, and transmitting the response reasoning content to the intelligent interaction terminal t