Search engine user intention recognition method based on multi-agent collaboration

Abstract

The invention discloses a search engine user intention recognition method based on multi-agent collaboration. The search engine user intention recognition method comprises the following steps: S1, constructing a state vector and generating a session pulse sequence; s2, constructing a multi-agent liquid state machine set, initializing parameters and establishing a communication channel; s3, inputting the pulse sequence into a session understanding state machine, outputting a semantic slot vector, and combining state vector scoring and searching to generate a candidate intention; s4, inputting the candidate intention and the semantic slot vector to a task planning state machine, constructing an intention graph and generating an evidence list; s5, inputting the state vector and the evidence list to an evidence retrieval state machine to obtain an evidence set; s6, inputting a state vector to a knowledge state machine to generate prior bias, and combining evidence and intention to perform aggregation; and S7, outputting an intention label according to a gating mechanism and updating parameters. According to the method, high-precision intention recognition in a multi-round search scene is realized, and the timeliness and credibility of a retrieval result are improved.

Description

Technical Field

[0001] This invention relates to the field of user intent recognition technology, and in particular to a method for user intent recognition in a search engine based on multi-agent collaboration. Background Technology

[0002] Existing search engine user intent recognition technologies mainly rely on single deep learning models for semantic understanding and intent classification. Common methods include classification structures based on recurrent neural networks, convolutional neural networks, and pre-trained language models. These methods can accomplish text representation and intent recognition tasks to a certain extent, but they often suffer from performance degradation when dealing with long dialogues, multi-turn interactions, and cross-domain intent recognition. Because semantic information has temporal dependencies during dialogue, traditional methods struggle to capture dynamic changes at different time scales during modeling, resulting in a lack of robustness and stability in intent recognition results.

[0003] Existing technologies also have shortcomings in evidence fusion and conflict resolution. Most methods simply weight the probability distribution of search results without modeling the contradictory relationships between evidence fragments. When data from different sources conflict in terms of theme or semantics, existing models are unable to make effective judgments and are prone to misjudgment. In real search scenarios, users often have personalized preferences and historical behaviors, but traditional methods lack full utilization of knowledge graphs and historical interaction data and fail to introduce prior bias information into the model, resulting in a gap between the intent recognition results and the user's actual needs.

[0004] In the multi-source retrieval and candidate fragment processing stages, most existing technologies rely on a single retrieval method and lack the combined use of rewritten queries, vector indexes, and click graph retrieval. As a result, they have limitations in coverage and accuracy. For candidate fragment screening, traditional methods lack systematic constraints in terms of time sensitivity, source credibility, and topic consistency control. Outdated information or low-quality data is easily mixed into the retrieval results. Due to the lack of multi-agent collaborative mechanisms and the ability of liquid state machines to model complex trajectories, existing technologies are not adaptable enough to dynamic environments and cannot effectively support user intent recognition in high-complexity search scenarios.

[0005] Therefore, how to provide a search engine user intent recognition method based on multi-agent collaboration is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a search engine user intent recognition method based on multi-agent collaboration. This invention introduces a liquid state machine structure and combines multi-agent collaborative processing mechanisms such as session understanding, task planning, evidence retrieval, knowledge modeling, evidence aggregation and arbitration. It describes in detail the complete process from query preprocessing to intent recognition, and has the advantages of strong temporal modeling capability, accurate conflict handling and high reliability of intent recognition results.

[0007] A method for recognizing user intent in a search engine based on multi-agent collaboration according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect user query requests and session context data, and preprocess them to obtain a session pulse sequence;

[0009] S2. Construct a multi-agent liquid state machine set, initialize the liquid pool parameters and establish a message channel;

[0010] S3. Input the conversation pulse sequence into the conversation understanding liquid state machine, run the liquid pool state propagation with a fixed step size, perform sequence labeling through the readout layer, output semantic slot vectors, perform joint scoring and sequence bundle search on the semantic slot vectors and conversation state vectors, and generate a candidate intent set based on the intent tag vocabulary and threshold rules.

[0011] S4. Input the candidate intent set and semantic slot vector into the task planning liquid state machine, construct an intent decomposition graph containing necessary nodes and alternative nodes based on the similarity of state trajectories, and map them to evidence types and field requirements according to the node dependency order to form an evidence requirement list.

[0012] S5. Input the session state vector and the evidence requirement list into the evidence retrieval liquid state machine to generate an evidence set;

[0013] S6. Input the session state vector into the knowledge liquid state machine to generate the prior bias vector. Input the evidence set, the prior bias vector and the candidate intent set into the evidence aggregation liquid state machine, calculate the intent confidence vector and the conflict score matrix, filter out mutually exclusive evidence according to the conflict threshold and generate the aggregation result data.

[0014] S7. Input the aggregated result data into the rule-gated readout layer of the arbitration liquid state machine, output the target intent label and structured parameters according to the hard gating rules and soft gating weights, update the liquid pool parameters in the multi-agent liquid state machine set, and complete the intent recognition output and online parameter update.

[0015] Optionally, the preprocessing in step S1 specifically includes: performing word segmentation, referential resolution, and time expression normalization on the user query request and session context data, concatenating them to generate a session state vector, and converting the session state vector into a pulse sequence using time arrival encoding and amplitude normalization.

[0016] Optionally, the multi-agent liquid state machine set in step S2 specifically includes: a conversation understanding liquid state machine, a task planning liquid state machine, an evidence retrieval liquid state machine, a knowledge liquid state machine, an evidence aggregation liquid state machine, and an arbitration liquid state machine.

[0017] Optionally, the step of initializing the liquid pool parameters and establishing the message channel in step S2 includes: setting the number of neurons, time constant range, synaptic sparsity, initial distribution of connection weights, conduction delay matrix and refractory period length for each liquid state machine in the multi-agent liquid state machine set; establishing a unidirectional message channel topology and specifying the message load field.

[0018] Optionally, step S3 specifically includes:

[0019] S31. Input the conversation pulse sequence into the conversation understanding liquid state machine, perform membrane potential update on the liquid pool neurons within a fixed step size, output a pulse and enter the refractory period when the membrane potential exceeds the threshold, and record the membrane potential value of the neuron at each step size and the pulse firing result in chronological order to form the liquid pool state sequence.

[0020] S32. Input the liquid pool state sequence into the readout layer. At each time step, the readout layer performs a weighted summation of the liquid pool state, combines it with the position embedding vector to classify the sequence, maps the classification result to the slot category vector, and concatenates them in time order to form a semantic slot vector.

[0021] S33. Jointly score the semantic slot vector and the session state vector, calculate the element-wise similarity between the semantic slot vector and the session state vector, introduce the context weight factor in the similarity calculation, give higher weight to the key slot positions, and form a joint scoring matrix.

[0022] S34. Perform sequence bundle search on the joint scoring matrix. First, set the search width and expand the candidate paths forward from each slot position in chronological order. Keep the first few paths with the highest cumulative scores. In the expansion process, introduce an intent label vocabulary to constrain the paths. When the slot label in a candidate path is inconsistent with the label in the vocabulary, discard the path. Several candidate paths that meet the vocabulary constraints are obtained.

[0023] S35. Filter out paths with confidence greater than the threshold from the candidate path set according to the set threshold, merge the slot labels in the path to form a high-confidence semantic slot vector.

[0024] S36. Perform label matching on the high-confidence semantic slot vectors based on the intent label vocabulary, compare the semantic slots with the predefined category items in the label vocabulary one by one, and fill the parameter values ​​with the parameter constraint fields to generate a candidate intent set.

[0025] Optionally, step S4 specifically includes:

[0026] S41. The candidate intent set and semantic slot vector are encoded into a task planning input pulse sequence using time arrival encoding and amplitude normalization. The input pulse sequence is then fed into the task planning liquid state machine. The state of the liquid pool neurons is advanced within a fixed step size, and the membrane potential and pulse firing results are recorded to generate the task planning liquid pool state sequence.

[0027] S42. Align the position of the liquid pool state sequence of the task planning, calculate the trajectory score corresponding to the tank in time order, remove the time slice with low score, retain the high score segment, and form a set of key trajectory segments.

[0028] S43. Based on the label categories and parameter constraints in the set of key trajectory segments and the set of candidate intentions, mark nodes that meet the conditions as required nodes and nodes with substitution relationships as replaceable nodes, and generate a draft intention decomposition graph containing nodes and their dependencies according to the order of dependencies between nodes.

[0029] S44. Perform a consistency check on the draft intent decomposition graph, remove nodes that are inconsistent with the semantic slot vector according to the node dependency order, merge duplicate nodes and eliminate conflict dependencies, and output the intent decomposition graph.

[0030] S45. According to the node dependency order and label category in the intent decomposition graph, retrieve the corresponding evidence type and field requirements in the intent label thesaurus, generate evidence requirement entries for each node, and the evidence requirement entries include evidence type, field name, field value range, time freshness requirements, source credibility requirements and topic consistency requirements, and accumulate to form an evidence requirement list.

[0031] Optionally, step S5 specifically includes:

[0032] Receive the session state vector and the evidence requirement list. For each evidence requirement item in the evidence requirement list, read the evidence type, field name, field value range, time freshness requirement, source credibility requirement and topic consistency requirement. Generate keyword sequence and time expression form based on field name and field value range, and combine them to form a rewrite query set.

[0033] In the lexical inverted index, the rewritten query set is used as the retrieval input to locate the document identifier and the start and end positions of the fragment containing the keyword sequence, extract the text fragment containing the field name and record the source identifier and timestamp to generate the first candidate fragment set;

[0034] Encode the session state vector into a dense vector, perform vector index retrieval using the nearest neighbor index, return the candidate fragment identifier that is closest to the dense vector, backfill the corresponding fragment content, source identifier and timestamp, and generate a second candidate fragment set;

[0035] In the historical click directed graph, a query node is established based on the session state vector. The walk step size and dwell probability are set, and the target type is limited to the evidence types in the evidence requirement list. A session consistency walk is performed on the query node. The result nodes with high access frequency are counted and the corresponding fragment content, source identifier and timestamp are extracted to generate a third candidate fragment set.

[0036] Merge the first candidate segment set, the second candidate segment set, and the third candidate segment set; perform deduplication based on the consistency of the source identifier and the start and end positions of the segments; and output the deduplicated candidate segment set.

[0037] For the deduplicated candidate fragment set, calculate the source weight, time weight, and topic consistency weight. The source weight is assigned a weight level according to the source credibility requirement, the time weight is assigned in descending order of the timestamp according to the time freshness requirement, and the topic consistency weight is assigned according to the similarity score returned by the nearest neighbor index. The weights are accumulated in order to obtain the comprehensive score.

[0038] Candidate fragments are grouped according to the evidence requirement items in the evidence requirement list. Each group is sorted in descending order based on the comprehensive score. The fragments with the highest ranking are selected to generate the results of the evidence requirement items. All results are then combined to form an evidence set.

[0039] Optionally, step S6 specifically includes:

[0040] S61. Input the session state vector into the knowledge liquid state machine, calculate the occurrence frequency and association strength based on the matching results of the liquid pool state sequence with the knowledge graph entities and historical interaction data, and normalize to generate the prior bias vector.

[0041] S62. Input the evidence set, prior bias vector and candidate intent set into the evidence aggregation liquid state machine, align the evidence fragments and intent labels with trajectories, calculate the corresponding probabilities and multiply and accumulate them with the prior bias weights to generate the intent confidence vector.

[0042] S63. Perform semantic comparison on the evidence set fragments and construct a conflict scoring matrix;

[0043] S64. Based on the conflict threshold, mutually exclusive segments are filtered out, and the confidence of the retained segments is normalized and weighted to output the aggregated result data.

[0044] Optionally, the rule-gated readout layer in step S7 specifically includes:

[0045] S71. Filter the aggregated result data according to the hard gating rules. The hard gating rules use business logic constraints and semantic consistency constraints as the judgment conditions. Aggregated results that do not conform to the preset intent category or do not meet the necessary slot integrity are directly removed, and candidate results that conform to the rules are retained.

[0046] S72. Apply soft-gating weights to the retained candidate results. The soft-gating weights are jointly determined by the historical interactive feedback data and the prior bias vector.

[0047] S73. Assign soft gating weights to the overall confidence of candidate results, perform weighted correction on the overall confidence, generate a weighted confidence distribution, select target intent labels based on the weighted confidence distribution, output the corresponding structured parameters, and compare the output with historical interactive feedback data to update the liquid pool parameters in the arbitration liquid state machine.

[0048] The beneficial effects of this invention are:

[0049] (1) This invention introduces a multi-agent liquid state machine set and uses a collaborative processing method of conversation understanding, task planning, evidence retrieval, knowledge modeling, evidence aggregation and arbitration to realize dynamic modeling of multi-turn dialogue and cross-domain search scenarios, solves the problem of decreased recognition accuracy of existing methods under long-term dependence, and improves the stability and robustness of user intent recognition.

[0050] (2) In the evidence fusion stage, the present invention introduces an intent confidence vector and a conflict score matrix, filters out mutually exclusive evidence by conflict threshold and performs weighted aggregation, ensures the semantic consistency and reliability of retrieval results from different sources, overcomes the shortcomings of existing technologies that cannot effectively handle evidence conflicts, and thus improves the consistency between the recognition results and the user's real needs.

[0051] (3) In the retrieval process, this invention combines rewritten query, vector index and click graph retrieval, and introduces multi-dimensional constraints such as time freshness, source credibility and topic consistency in the candidate fragment selection, so that the retrieval results are optimized in terms of timeliness, credibility and relevance, and the accuracy and practical value of the final intent recognition results are improved. Attached Figure Description

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0053] Figure 1 This is an overall flowchart of a search engine user intent recognition method based on multi-agent collaboration proposed in this invention;

[0054] Figure 2This is a flowchart of evidence retrieval and candidate fragment weighted filtering for a search engine user intent recognition method based on multi-agent collaboration proposed in this invention.

[0055] Figure 3 This is a structured processing flowchart for evidence aggregation and arbitration intent output of a search engine user intent recognition method based on multi-agent collaboration proposed in this invention. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0057] refer to Figure 1-3 A method for recognizing user intent in a search engine based on multi-agent collaboration includes the following steps:

[0058] S1. Collect user query requests and session context data, construct a session state vector, and preprocess to obtain a session pulse sequence;

[0059] S2. Construct a multi-agent liquid state machine set, initialize the liquid pool parameters and establish a message channel;

[0060] S3. Input the conversation pulse sequence into the conversation understanding liquid state machine, run the liquid pool state propagation with a fixed step size, perform sequence labeling through the readout layer, output semantic slot vectors, perform joint scoring and sequence bundle search on the semantic slot vectors and conversation state vectors, and generate a candidate intent set based on the intent tag vocabulary and threshold rules.

[0061] S4. Input the candidate intent set and semantic slot vector into the task planning liquid state machine, construct an intent decomposition graph containing necessary nodes and alternative nodes based on the similarity of state trajectories, and map them to evidence types and field requirements according to the node dependency order to form an evidence requirement list.

[0062] S5. Input the session state vector and evidence requirement list into the evidence retrieval liquid state machine, execute query rewriting retrieval to generate rewritten queries and retrieve candidate fragments in the lexical inverted index, execute vector index retrieval to encode the session state vector into a dense vector and retrieve candidate fragments in the nearest neighbor index, execute click graph retrieval to obtain candidate fragments by performing session consistency walk in the historical click directed graph, and perform deduplication, source credibility weighting, time freshness weighting and topic consistency weighting on the three types of candidate fragments to generate an evidence set;

[0063] S6. Input the session state vector into the knowledge liquid state machine to generate the prior bias vector. Input the evidence set, the prior bias vector and the candidate intent set into the evidence aggregation liquid state machine, calculate the intent confidence vector and the conflict score matrix, filter out mutually exclusive evidence according to the conflict threshold and generate the aggregation result data.

[0064] S7. Input the aggregated result data into the rule-gated readout layer of the arbitration liquid state machine, output the target intent label and structured parameters according to the hard gating rules and soft gating weights, update the liquid pool parameters in the multi-agent liquid state machine set, and complete the intent recognition output and online parameter update.

[0065] In one embodiment of the present invention, after receiving the aggregation result data, the arbitration liquid state machine uses a rule-gated readout layer to output the target intent label and structured parameters. The hard-gating rules strictly screen out candidate results that do not meet the conditions based on semantic consistency and slot integrity. The soft-gating weights adjust the confidence of the remaining candidate results based on historical interaction feedback data and prior bias vectors to obtain the corrected intent label and parameters.

[0066] After completing the target intent labeling and structured parameter output, the arbitration liquid state machine compares the actual output with user feedback data, including click behavior, dwell time, and subsequent query operations. Based on the comparison results, the system dynamically updates the liquid pool parameters in the multi-agent liquid state machine set. The updates include adjusting the time constant parameter to improve the neurons' ability to remember long-term dependencies, correcting the conduction delay matrix to optimize the efficiency of information transmission between neurons, and fine-tuning the refractory period length to balance frequent triggering and stability.

[0067] The updated liquid pool parameters are directly applied in the next round of session processing, enabling multi-agents to converge faster and improve the accuracy of intent recognition when processing similar inputs. In this way, the intent recognition results are not only optimized in the current session, but also show continuous improvement in subsequent sessions, realizing an online adaptive parameter update mechanism.

[0068] For example, in a continuous session, the user inputs the query "new energy policy". The system outputs the target intent label "new energy industry policy and regulations" and the structured parameter "latest policy documents" in the arbitration liquid state machine. The user then inputs the query "subsidy standards". The feedback data shows that the user's click and reading time is concentrated on subsidy-related content, while the time spent on policy document content is shorter. Based on this feedback result, the system updates the liquid pool parameters of the multi-agent liquid state machine set, adjusting the time constant parameter to a longer time series range to enhance the model's ability to remember cross query contexts, while shortening the refractory period to improve the response sensitivity to rapid continuous input.

[0069] With the updated parameters, the system can directly identify the user's intent in the next round of dialogue as "details of new energy subsidy policy" when inputting "conditions for enterprises to receive subsidies," and accurately output the required structured fields. Compared with the model before the update, the intent recognition accuracy improved from 87.5% to 92.3%, and the user's satisfaction score after this round of query increased to 4.7 points. This example shows that by combining intent recognition output with online parameter updates, the system can continuously optimize based on actual interaction performance, ensuring the accuracy and stability of recognition results in subsequent sessions.

[0070] In this embodiment, the preprocessing in step S1 specifically includes: performing word segmentation, referential resolution, and time expression normalization on the user query request and session context data, concatenating them to generate a session state vector, and converting the session state vector into a pulse sequence using time arrival encoding and amplitude normalization.

[0071] In this embodiment, the multi-agent liquid state machine set in step S2 specifically includes: conversation understanding liquid state machine, task planning liquid state machine, evidence retrieval liquid state machine, knowledge liquid state machine, evidence aggregation liquid state machine, and arbitration liquid state machine.

[0072] In this embodiment, the step of initializing the liquid pool parameters and establishing the message channel in step S2 includes: setting the number of neurons, time constant range, synaptic sparsity, connection weight initialization distribution, conduction delay matrix and refractory period length for each liquid state machine in the multi-agent liquid state machine set; establishing a unidirectional message channel topology and specifying the message load field.

[0073] In this embodiment, step S3 specifically includes:

[0074] S31. Input the conversation pulse sequence into the conversation understanding liquid state machine, and perform membrane potential update on the liquid pool neurons within a fixed step size. Specifically, receive the input pulse and filter effective connections according to synaptic sparsity. Accumulate the effective connection pulses with connection weights. Delay the pulses by several steps according to the conduction delay matrix and superimpose them onto the input. Combine the time constant parameter to exponentially decay the residual potential of the previous moment and superimpose it with the current synaptic current to obtain the updated membrane potential. When the membrane potential exceeds the threshold, output a pulse and enter the refractory period. Record the membrane potential values ​​of the neurons at each step size and the pulse firing results in chronological order to form the liquid pool state sequence.

[0075] S32. Input the liquid pool state sequence into the readout layer. At each time step, the readout layer performs a weighted summation of the liquid pool state, combines it with the position embedding vector to classify the sequence, maps the classification result to the slot category vector, and concatenates them in time order to form a semantic slot vector.

[0076] S33. Jointly score the semantic slot vector and the session state vector, calculate the element-wise similarity between the semantic slot vector and the session state vector, introduce the context weight factor in the similarity calculation, give higher weight to the key slot positions, and form a joint scoring matrix.

[0077] S34. Perform sequence bundle search on the joint scoring matrix. First, set the search width and expand the candidate paths forward from each slot position in chronological order. Keep the first few paths with the highest cumulative scores. In the expansion process, introduce an intent label vocabulary to constrain the paths. When the slot label in a candidate path is inconsistent with the label in the vocabulary, discard the path. Several candidate paths that meet the vocabulary constraints are obtained.

[0078] S35. Filter out paths with confidence greater than the threshold from the candidate path set according to the set threshold, merge the slot labels in the path to form a high-confidence semantic slot vector.

[0079] S36. Perform label matching on the high-confidence semantic slot vectors based on the intent label vocabulary, compare the semantic slots with the predefined category items in the label vocabulary one by one, and fill the parameter values ​​with the parameter constraint fields to generate a candidate intent set.

[0080] In this embodiment, step S4 specifically includes:

[0081] S41. The candidate intent set and semantic slot vector are encoded into a task planning input pulse sequence using time arrival encoding and amplitude normalization. The input pulse sequence is then fed into the task planning liquid state machine. The state of the liquid pool neurons is advanced within a fixed step size, and the membrane potential and pulse firing results are recorded to generate the task planning liquid pool state sequence.

[0082] S42. Align the positional sequence of the liquid pool state in the task planning, calculate the trajectory score corresponding to the tank position in chronological order, remove low-scoring time slices, and retain high-scoring segments to form a set of key trajectory segments; specifically:

[0083] Based on the slot label order and parameter constraints in the semantic slot vector, the trajectory of the liquid pool state sequence in the task planning is scored and segments are selected in chronological order. Specifically:

[0084] Construct a time window, set a fixed window length W and a step interval Δt, and perform sliding segmentation on the liquid pool state sequence of the task planning to obtain a window sequence sorted by time;

[0085] Within each time window, calculate the alignment consistency tag sequence, match the slot label corresponding to the center of the window with the slot label at the same position in the semantic slot vector, mark consistency as 1, mark inconsistency as 0, and obtain the time series of consistency tags;

[0086] Within each time window, the pulse activity density is calculated. The sum of the number of pulses fired by all neurons within the window is divided by the total number of steps that can be fired within the window and the number of neurons to obtain an activity density value between 0 and 1.

[0087] Calculate the membrane potential stabilization ratio within each time window, and make a threshold judgment on the membrane potential difference between adjacent steps. The proportion of steps with a difference not exceeding the stabilization threshold to the number of steps in the window is taken as the stabilization ratio.

[0088] The consistency marker, activity density, and stability ratio are weighted and summed according to preset weights to obtain the comprehensive score sequence for this time window. The comprehensive score is then normalized to the range of 0 to 1.

[0089] Set a low threshold T_low, a segment threshold T_seg, a retention threshold T_keep, and a minimum segment length L_min. Connect windows with a comprehensive score not lower than T_seg in a continuous time window to form candidate segments. Delete candidate segments with a length less than L_min. Delete candidate segments with a mean comprehensive score lower than T_low. Retain candidate segments with a mean comprehensive score not lower than T_keep to form a set of key trajectory segments.

[0090] S43. Based on the label categories and parameter constraints in the set of key trajectory segments and the set of candidate intentions, mark nodes that meet the conditions as required nodes and nodes with substitution relationships as replaceable nodes, and generate a draft intention decomposition graph containing nodes and their dependencies according to the order of dependencies between nodes.

[0091] S44. Perform a consistency check on the draft intent decomposition graph, remove nodes that are inconsistent with the semantic slot vector according to the node dependency order, merge duplicate nodes and eliminate conflict dependencies, and output the intent decomposition graph.

[0092] S45. According to the node dependency order and label category in the intent decomposition graph, retrieve the corresponding evidence type and field requirements from the intent label thesaurus, and generate an evidence requirement entry for each node. Each evidence requirement entry includes the evidence type, field name, field value range, time freshness requirement, source credibility requirement, and topic consistency requirement, accumulating to form an evidence requirement list. Among these:

[0093] The time freshness requirement is used to ensure that search results reflect the latest context and real-time status of the user's query. Specifically, when generating evidence requirement entries, a valid time range is pre-defined for different intent categories, such as "news data within the last hour," "transaction data within the last seven days," or "comment data within the last month." When scoring candidate segments, the timestamp is compared with the preset range. Segments within the range are given high weight, while segments outside the range have their weight reduced using a decreasing function. This makes the output evidence set more in line with the user's implicit need for timeliness.

[0094] Source credibility requirements are used to ensure that search results come from credible data sources. Specifically, during system construction, a credibility level is assigned to each data source, such as three levels: "authoritative source," "verified source," and "user-generated content," each corresponding to a different credibility score. When scoring candidate fragments, fragments from authoritative sources are given higher weights, while non-verified fragments generated solely by users are given lower weights, thereby suppressing the impact of noise and low-quality data on intent recognition results.

[0095] The topic consistency requirement ensures that the content of the retrieved fragments aligns with the topic expressed by the session state vector. Specifically, during vector index retrieval, a similarity score is calculated between the candidate fragment dense vector and the session state vector. A higher similarity score indicates greater consistency between the candidate fragment and the user's query topic. When scoring candidate fragments, the normalized similarity score is used as a weighted average for topic consistency. If the similarity score is below a set threshold, the fragment is directly removed, ensuring that the final evidence set contains only fragments highly relevant to the user's query semantics.

[0096] In this embodiment, step S5 specifically includes:

[0097] Receive the session state vector and the evidence requirement list. For each evidence requirement item in the evidence requirement list, read the evidence type, field name, field value range, time freshness requirement, source credibility requirement and topic consistency requirement. Generate keyword sequence and time expression form based on field name and field value range, and combine them to form a rewrite query set.

[0098] In the lexical inverted index, the rewritten query set is used as the retrieval input to locate the document identifier and the start and end positions of the fragment containing the keyword sequence, extract the text fragment containing the field name and record the source identifier and timestamp to generate the first candidate fragment set;

[0099] Encode the session state vector into a dense vector, perform vector index retrieval using the nearest neighbor index, return the candidate fragment identifier that is closest to the dense vector, backfill the corresponding fragment content, source identifier and timestamp, and generate a second candidate fragment set;

[0100] In the historical click directed graph, a query node is established based on the session state vector. The walk step size and dwell probability are set, and the target type is limited to the evidence types in the evidence requirement list. A session consistency walk is performed on the query node. The result nodes with high access frequency are counted and the corresponding fragment content, source identifier and timestamp are extracted to generate a third candidate fragment set.

[0101] Merge the first candidate segment set, the second candidate segment set, and the third candidate segment set; perform deduplication based on the consistency of the source identifier and the start and end positions of the segments; and output the deduplicated candidate segment set.

[0102] For the deduplicated candidate fragment set, calculate the source weight, time weight, and topic consistency weight. The source weight is assigned a weight level according to the source credibility requirement, the time weight is assigned in descending order of the timestamp according to the time freshness requirement, and the topic consistency weight is assigned according to the similarity score returned by the nearest neighbor index. The weights are accumulated in order to obtain the comprehensive score.

[0103] Candidate fragments are grouped according to the evidence requirement items in the evidence requirement list. Each group is sorted in descending order based on the comprehensive score. The fragments with the highest ranking are selected to generate the results of the evidence requirement items. All results are then combined to form an evidence set.

[0104] In this embodiment, step S6 specifically includes:

[0105] S61. Input the session state vector into the knowledge liquid state machine, calculate the occurrence frequency and association strength based on the matching results of the liquid pool state sequence with the knowledge graph entities and historical interaction data, and normalize to generate the prior bias vector.

[0106] Specifically, the process of generating the prior bias vector includes:

[0107] First, the session state vector is input into the knowledge liquid state machine. The neuron state is advanced through the liquid pool at a fixed step size. The resulting liquid pool state sequence is used to characterize the dynamic semantic features of the session.

[0108] The system extracts a set of entities related to the session state from the knowledge graph, performs similarity matching between the feature fragments in the liquid pool state sequence and the embedding vectors of the entities in the knowledge graph, and combines the query records and click behaviors in the user's historical interaction data to calculate the frequency of each entity in the current session and the strength of its association with historical behaviors.

[0109] The occurrence frequency and association strength are weighted and accumulated to form an entity weight distribution. After normalization, the entity weight distribution is transformed into a numerically stable prior bias vector.

[0110] The prior bias vector is used to weight different evidence fragments in the evidence aggregation liquid state machine, thereby guiding the model to pay more attention to candidate intents that are strongly related to the user's historical behavior and knowledge graph when calculating intent confidence.

[0111] S62. Input the evidence set, prior bias vector and candidate intent set into the evidence aggregation liquid state machine, align the evidence fragments and intent labels with trajectories, calculate the corresponding probabilities and multiply and accumulate them with the prior bias weights to generate the intent confidence vector.

[0112] Specifically, the trajectory alignment process is as follows:

[0113] Each piece of evidence in the evidence set is encoded as a pulse input sequence and input into the evidence aggregation liquid state machine. It is injected into the liquid pool synchronously with the pulse input sequence of the intent label in the candidate intent set. The liquid pool neurons advance the state within a fixed step size, forming the corresponding evidence trajectory and intent trajectory.

[0114] In the liquid pool state sequence, the evidence trajectory and the intention trajectory are compared step by step with time step as index. The neuronal pulse firing pattern and membrane potential change trend are extracted, and the trajectory similarity between the two in the same time window is calculated.

[0115] The trajectory similarity sequence is accumulated to obtain the corresponding probability between each evidence fragment and the target intent label. The corresponding probability is multiplied by the weight value in the prior bias vector and accumulated in the evidence fragment dimension to generate the intent confidence vector, which provides input for subsequent aggregation and arbitration.

[0116] S63. Perform semantic comparison on the evidence set fragments and construct a conflict scoring matrix. Specifically, the fragments in the evidence set are encoded as semantic vectors. By calculating the semantic similarity between fragments and the similarity of the liquid pool trajectory, the semantic difference between each pair of fragments is obtained. When the difference exceeds a set threshold, it is determined that there is a conflict relationship between the fragments. The conflict intensity is recorded in a matrix with evidence fragments as rows and candidate intent labels as columns to form a conflict scoring matrix.

[0117] S64. Based on the conflict threshold, mutually exclusive segments are filtered out, and the confidence of the retained segments is normalized and weighted to output the aggregated result data.

[0118] Specifically, the process of generating the aggregation result data is as follows:

[0119] After constructing the conflict scoring matrix, the evidence aggregation liquid state machine compares the conflict intensity of each pair of evidence fragments in the candidate intent label dimension with a preset conflict threshold. If the conflict intensity is greater than the threshold, the fragment pair is determined to be mutually exclusive, and the fragment with the lower score is removed; if the conflict intensity is less than or equal to the threshold, the fragment is retained.

[0120] For the retained set of segments, the system reads the corresponding intent confidence vector, normalizes the confidence values ​​of each segment, and distributes all confidence values ​​within the range of [0,1] to ensure that different segments are comparable at the same scale.

[0121] Based on the normalized confidence scores, a weighted summation is performed on the segments under each candidate intent tag to obtain the comprehensive confidence score of that intent tag, and the segments are then sorted according to the comprehensive confidence scores.

[0122] The final output aggregated result data consists of the content of the retained fragment, source identifier, timestamp, and its comprehensive confidence level. This aggregated result data is passed as input to the arbitration liquid state machine.

[0123] In this embodiment, the rule-gated readout layer in step S7 specifically includes:

[0124] S71. Filter the aggregated result data according to the hard gating rules. The hard gating rules use business logic constraints and semantic consistency constraints as the judgment conditions. Aggregated results that do not conform to the preset intent category or do not meet the necessary slot integrity are directly removed, and candidate results that conform to the rules are retained.

[0125] S72. Apply soft-gating weights to the retained candidate results. The soft-gating weights are jointly determined by historical interaction feedback data and prior bias vectors. Historical interaction feedback data is used to reflect the statistical characteristics of user click behavior and dwell time, and prior bias vectors are used to reflect the association strength between knowledge graph entities and session context.

[0126] S73. Assign soft gating weights to the overall confidence of candidate results, perform weighted correction on the overall confidence, generate a weighted confidence distribution, select target intent labels based on the weighted confidence distribution, output the corresponding structured parameters, and compare the output with historical interactive feedback data to update the liquid pool parameters in the arbitration liquid state machine.

[0127] Example 1:

[0128] In a real-world application environment of a large search platform, users frequently input multiple rounds of queries, covering not only everyday life issues but also cross-disciplinary professional questions. For example, users might ask queries like "latest changes in new energy policies," "subsidy measures for corresponding companies," or "relevant legal clauses" in consecutive sessions. Traditional search engine intent recognition methods often struggle to maintain a coherent understanding of the semantic context in long-session scenarios, and their ability to identify cross-disciplinary questions is insufficient. Furthermore, the search results provided by the system may contain outdated information, web content from low-reliability sources, or even content whose topics are inconsistent with the query semantics, forcing users to repeatedly modify keywords and manually filter information, severely impacting efficiency.

[0129] In this embodiment of the invention, the user's query request and session context data are preprocessed to generate a session pulse sequence, which is then processed by a multi-agent liquid state machine set. Liquid state machines with different roles respectively undertake the functions of semantic slot extraction, intent candidate generation, task decomposition, evidence retrieval, knowledge bias generation, evidence aggregation, and arbitration decision-making. The session understanding liquid state machine labels the user's continuous input with slots and generates a candidate intent set using a bundle search strategy. The task planning liquid state machine maps the candidate intents to an intent decomposition graph and an evidence requirement list. The evidence retrieval liquid state machine then executes rewriting queries, vector indexing, and click graph retrieval in parallel, acquiring candidate fragments from multiple channels. During the candidate fragment fusion process, the system introduces weighted calculations based on three dimensions: time freshness, source credibility, and topic consistency, filtering out outdated, low-quality, and semantically biased data. The knowledge liquid state machine combines historical query behavior and knowledge graph information to generate a priori bias vector, guiding the evidence aggregation liquid state machine to perform conflict discrimination and confidence calculation. Finally, the rule-gated readout layer of the arbitration liquid state machine completes hard rule filtering and soft weight correction, outputting structured intent labels and parameters.

[0130] To verify the effectiveness of this invention in complex search tasks, it was compared with traditional deep learning classification models and single-channel retrieval models. Three typical scenarios were simulated in the experimental environment: news and information queries with high timeliness requirements, policy and regulation queries with high requirements for source authority, and comprehensive queries with cross-domain semantics. Thousands of real user conversation data were collected in each scenario, and a set of reference intent tags was established through manual annotation. The experiment compared the performance of the three methods in terms of intent recognition accuracy, intent recognition recall, timeliness of search results, credibility of search result sources, and consistency of search result topics.

[0131] Data results show that this invention outperforms the comparison methods in both intent recognition accuracy and recall, especially in long-conversation cross-domain tasks, where the accuracy of traditional models drops to around 72%, while this invention maintains above 90%. In news queries, under the influence of time freshness weighting, this invention returns over 95% of data from within the past week, compared to only 64% for traditional methods. In policy and regulation queries, the proportion of results with a source credibility score above 0.85 reaches 92%, while the traditional method achieves 70%. In comprehensive query tasks, the proportion of segments with a topic consistency score exceeding 0.8 reaches 89%, while the traditional method achieves only 67%. These data fully demonstrate that this invention, by introducing a multi-agent liquid state machine and a multi-dimensional constraint mechanism, can better capture temporal features, resolve cross-domain conflicts, and ensure the freshness and credibility of results. The table below shows the comparative data of different methods in typical search tasks.

[0132] Table 1: Performance Comparison Data of Different Methods in Search Tasks

[0133] Indicator Categories Method of the present invention Traditional deep learning classification models Single-channel retrieval model Intent recognition accuracy 91.2％ 78.5％ 74.3％ Intent recognition recall rate 93.6％ 80.1％ 76.8％ Search timeliness 95.4％ 64.2％ 61.9％ Source credibility 92.1％ 70.3％ 66.7％ Theme Consistency 89.0％ 67.4％ 62.5％ Average response time / s 1.35 1.42 1.28 Conflict Fragment Removal Rate 87.8％ 58.6％ 55.1％ User satisfaction rating 4.6 / 5 3.8 / 5 3.5 / 5

[0134] As can be seen from the table results, the method of this invention exhibits significant advantages in all key performance indicators. The intent recognition accuracy and recall rate reach 91.2% and 93.6% respectively, which are significantly improved compared to the 78.5% and 80.1% of traditional deep learning classification models, and even more significantly compared to the 74.3% and 76.8% of single-channel retrieval models. This indicates that the present invention has stronger stability and robustness in complex conversational semantic understanding and cross-domain intent recognition. In terms of retrieval timeliness, the method of this invention can ensure that 95.4% of the results come from the latest data, while the proportion of traditional deep learning models and single-channel models is less than 65%, reflecting the advantage of the present invention under the constraint of time freshness.

[0135] Regarding source credibility, 92.1% of the results output by the method of this invention come from authoritative and reliable data sources, while the traditional methods perform at 70.3% and 66.7% respectively, indicating that the multidimensional constraint mechanism can effectively suppress low-quality sources. In terms of topic consistency, 89.0% of the results of this invention are highly consistent with the user's query topic, while the comparative methods are 67.4% and 62.5% respectively, demonstrating the important role of semantic consistency control. The conflict fragment removal rate reaches 87.8%, which is significantly higher than the less than 60% level of the traditional method, proving that the conflict scoring matrix and threshold screening mechanism effectively improve the quality of evidence fusion.

[0136] Although the average response time is slightly increased to 1.35 seconds, which is slightly higher than the 1.28 seconds of the single-channel retrieval model, this difference is acceptable while ensuring higher accuracy and reliability. Meanwhile, the user satisfaction score reaches 4.6, significantly higher than the 3.8 and 3.5 of traditional methods, reflecting user approval of the quality of search results and the accuracy of intent recognition. Overall, this invention, while ensuring efficient response, improves the timeliness, reliability, and consistency of search engine intent recognition, meeting user needs in complex search scenarios.

[0137] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for recognizing user intent in a search engine based on multi-agent collaboration, characterized in that, Includes the following steps: S1. Collect user query requests and session context data, construct a session state vector, and preprocess to obtain a session pulse sequence; S2. Construct a multi-agent liquid state machine set, initialize the liquid pool parameters and establish a message channel; S3. Input the conversation pulse sequence into the conversation understanding liquid state machine, run the liquid pool state propagation with a fixed step size, perform sequence labeling through the readout layer, output semantic slot vectors, perform joint scoring and sequence bundle search on the semantic slot vectors and conversation state vectors, and generate a candidate intent set based on the intent tag vocabulary and threshold rules. S4. Input the candidate intent set and semantic slot vector into the task planning liquid state machine, construct an intent decomposition graph containing necessary nodes and alternative nodes based on the similarity of state trajectories, and map them to evidence types and field requirements according to the node dependency order to form an evidence requirement list. S5. Input the session state vector and the evidence requirement list into the evidence retrieval liquid state machine to generate an evidence set; S6. Input the session state vector into the knowledge liquid state machine to generate the prior bias vector. Input the evidence set, the prior bias vector and the candidate intent set into the evidence aggregation liquid state machine, calculate the intent confidence vector and the conflict score matrix, filter out mutually exclusive evidence according to the conflict threshold and generate the aggregation result data. S7. Input the aggregation result data into the rule-gated readout layer of the arbitration liquid state machine, update the liquid pool parameters in the multi-agent liquid state machine set, and complete the intent recognition output and online parameter update.

2. The method for recognizing user intent in a search engine based on multi-agent collaboration according to claim 1, characterized in that, The preprocessing in step S1 specifically includes: performing word segmentation, referential resolution, and time expression normalization on the user query request and session context data, concatenating them to generate a session state vector, and converting the session state vector into a pulse sequence using time arrival encoding and amplitude normalization.

3. The method for recognizing user intent in a search engine based on multi-agent collaboration according to claim 2, characterized in that, The multi-agent liquid state machine set in step S2 specifically includes: conversation understanding liquid state machine, task planning liquid state machine, evidence retrieval liquid state machine, knowledge liquid state machine, evidence aggregation liquid state machine, and arbitration liquid state machine.

4. The method for recognizing user intent in a search engine based on multi-agent collaboration according to claim 3, characterized in that, The steps of initializing the liquid pool parameters and establishing the message channel in step S2 include: setting the number of neurons, time constant range, synaptic sparsity, initial distribution of connection weights, conduction delay matrix and refractory period length for each liquid state machine in the multi-agent liquid state machine set; establishing a unidirectional message channel topology and specifying the message load field.

5. The method for recognizing user intent in a search engine based on multi-agent collaboration according to claim 4, characterized in that, Step S3 specifically includes: S31. Input the conversation pulse sequence into the conversation understanding liquid state machine, perform membrane potential update on the liquid pool neurons within a fixed step size, output a pulse and enter the refractory period when the membrane potential exceeds the threshold, and record the membrane potential value of the neuron at each step size and the pulse firing result in chronological order to form the liquid pool state sequence. S32. Input the liquid pool state sequence into the readout layer. At each time step, the readout layer performs a weighted summation of the liquid pool state, combines it with the position embedding vector to classify the sequence, maps the classification result to the slot category vector, and concatenates them in time order to form a semantic slot vector. S33. Jointly score the semantic slot vector and the session state vector, calculate the element-wise similarity between the semantic slot vector and the session state vector, introduce the context weight factor in the similarity calculation, give higher weight to the key slot positions, and form a joint scoring matrix. S34. Perform sequence bundle search on the joint scoring matrix. First, set the search width and expand the candidate paths forward from each slot position in chronological order. Keep the first few paths with the highest cumulative scores. In the expansion process, introduce an intent label vocabulary to constrain the paths. When the slot label in a candidate path is inconsistent with the label in the vocabulary, discard the path. Several candidate paths that meet the vocabulary constraints are obtained. S35. Filter out paths with confidence greater than the threshold from the candidate path set according to the set threshold, merge the slot labels in the path to form a high-confidence semantic slot vector. S36. Perform label matching on the high-confidence semantic slot vectors based on the intent label vocabulary, compare the semantic slots with the predefined category items in the label vocabulary one by one, and fill the parameter values ​​with the parameter constraint fields to generate a candidate intent set.

6. The method for recognizing user intent in a search engine based on multi-agent collaboration according to claim 5, characterized in that, Step S4 specifically includes: S41. The candidate intent set and semantic slot vector are encoded into a task planning input pulse sequence using time arrival encoding and amplitude normalization. The input pulse sequence is then fed into the task planning liquid state machine. The state of the liquid pool neurons is advanced within a fixed step size, and the membrane potential and pulse firing results are recorded to generate the task planning liquid pool state sequence. S42. Align the position of the liquid pool state sequence of the task planning, calculate the trajectory score corresponding to the tank in time order, remove the time slice with low score, retain the high score segment, and form a set of key trajectory segments. S43. Based on the label categories and parameter constraints in the set of key trajectory segments and the set of candidate intentions, mark nodes that meet the conditions as required nodes and nodes with substitution relationships as replaceable nodes, and generate a draft intention decomposition graph containing nodes and their dependencies according to the order of dependencies between nodes. S44. Perform a consistency check on the draft intent decomposition graph, remove nodes that are inconsistent with the semantic slot vector according to the node dependency order, merge duplicate nodes and eliminate conflict dependencies, and output the intent decomposition graph. S45. According to the node dependency order and label category in the intent decomposition graph, retrieve the corresponding evidence type and field requirements in the intent label thesaurus, generate evidence requirement entries for each node, and the evidence requirement entries include evidence type, field name, field value range, time freshness requirements, source credibility requirements and topic consistency requirements, and accumulate to form an evidence requirement list.

7. The method for recognizing user intent in a search engine based on multi-agent collaboration according to claim 6, characterized in that, Step S5 specifically includes: Receive the session state vector and the evidence requirement list. For each evidence requirement item in the evidence requirement list, read the evidence type, field name, field value range, time freshness requirement, source credibility requirement and topic consistency requirement. Generate keyword sequence and time expression form based on field name and field value range, and combine them to form a rewrite query set. In the lexical inverted index, the rewritten query set is used as the retrieval input to locate the document identifier and the start and end positions of the fragment containing the keyword sequence, extract the text fragment containing the field name and record the source identifier and timestamp to generate the first candidate fragment set; Encode the session state vector into a dense vector, perform vector index retrieval using the nearest neighbor index, return the candidate fragment identifier that is closest to the dense vector, backfill the corresponding fragment content, source identifier and timestamp, and generate a second candidate fragment set; In the historical click directed graph, a query node is established based on the session state vector. The walk step size and dwell probability are set, and the target type is limited to the evidence types in the evidence requirement list. A session consistency walk is performed on the query node. The result nodes with high access frequency are counted and the corresponding fragment content, source identifier and timestamp are extracted to generate a third candidate fragment set. Merge the first candidate segment set, the second candidate segment set, and the third candidate segment set; perform deduplication based on the consistency of the source identifier and the start and end positions of the segments; and output the deduplicated candidate segment set. For the deduplicated candidate fragment set, calculate the source weight, time weight, and topic consistency weight. The source weight is assigned a weight level according to the source credibility requirement, the time weight is assigned in descending order of the timestamp according to the time freshness requirement, and the topic consistency weight is assigned according to the similarity score returned by the nearest neighbor index. The weights are accumulated in order to obtain the comprehensive score. Candidate fragments are grouped according to the evidence requirement items in the evidence requirement list. Each group is sorted in descending order based on the comprehensive score. The fragments with the highest ranking are selected to generate the results of the evidence requirement items. All results are then combined to form an evidence set.

8. The method for recognizing user intent in a search engine based on multi-agent collaboration according to claim 7, characterized in that, Step S6 specifically includes: S61. Input the session state vector into the knowledge liquid state machine, calculate the occurrence frequency and association strength based on the matching results of the liquid pool state sequence with the knowledge graph entities and historical interaction data, and normalize to generate the prior bias vector. S62. Input the evidence set, prior bias vector and candidate intent set into the evidence aggregation liquid state machine, align the evidence fragments and intent labels with trajectories, calculate the corresponding probabilities and multiply and accumulate them with the prior bias weights to generate the intent confidence vector. S63. Perform semantic comparison on the evidence set fragments and construct a conflict scoring matrix; S64. Based on the conflict threshold, mutually exclusive segments are filtered out, and the confidence of the retained segments is normalized and weighted to output the aggregated result data.

9. A method for recognizing user intent in a search engine based on multi-agent collaboration as described in claim 8, characterized in that, The rule-gated readout layer in step S7 specifically includes: S71. Filter the aggregated result data according to the hard gating rules. The hard gating rules use business logic constraints and semantic consistency constraints as the judgment conditions. Aggregated results that do not conform to the preset intent category or do not meet the necessary slot integrity are directly removed, and candidate results that conform to the rules are retained. S72. Apply soft-gating weights to the retained candidate results. The soft-gating weights are jointly determined by the historical interactive feedback data and the prior bias vector. S73. Assign soft gating weights to the overall confidence of candidate results, perform weighted correction on the overall confidence, generate a weighted confidence distribution, select target intent labels based on the weighted confidence distribution, output the corresponding structured parameters, and compare the output with historical interactive feedback data to update the liquid pool parameters in the arbitration liquid state machine.

Images