A method and system for searching by combining user input information

Abstract

This invention discloses a method and system for searching by merging user input information, specifically relating to the field of language parsing technology. The method includes the following steps: First, multiple user inputs are transformed into an input semantic structure containing keywords, semantic relationships, and contextual labels; then, the interaction focus is extracted by comparing each structure, and a semantic consensus score is generated; a fusion path graph is constructed based on the consensus score, keyword weights, and semantic fit; semantic fusion and conflict resolution are performed according to the path graph to generate a unified expression structure; finally, keywords are sorted according to semantic level and task orientation to construct a comprehensive search expression to complete the query output. This invention extracts consensus content and constructs a fusion path graph through semantic parsing and cross-comparison of multiple user inputs, ensuring a reasonable fusion order and semantic consistency; finally, a comprehensive semantic expression structure is generated and a search expression is constructed, achieving precise alignment between query content and user intent, and improving the representativeness and accuracy of search results.

Description

Technical Field

[0001] This invention relates to the field of language parsing technology, and more specifically, to a method and system for performing searches by merging user input information. Background Technology

[0002] With the rapid development of artificial intelligence, natural language processing, and human-computer interaction technologies, various intelligent systems are increasingly incorporating natural language input from multiple users, particularly in collaborative office systems, group decision-making platforms, intelligent search assistants, and multi-user service terminals. In these practical application scenarios, systems often receive language content input from multiple users around the same task objective or query requirement. These inputs differ significantly in surface semantics, pragmatic context, vocabulary selection, and structural expression, exhibiting obvious heterogeneity, contextual discreteness, and diverse intents.

[0003] Under current technological conditions, mainstream search systems or semantic understanding frameworks are typically based on single-user input models for semantic parsing and query generation, with most processes focusing on keyword extraction, word frequency weighting, or phrase matching. Although recent research has attempted to introduce contextual embedding or user behavior features to enhance semantic models, they still fall short in handling simultaneous multi-user input.

[0004] Specifically, existing technologies generally suffer from the following technical bottlenecks:

[0005] When faced with multiple user inputs, the system often processes them equivalently or directly concatenates them, failing to construct a unified semantic structure to extract consensus. This leads to redundant expressions or overlapping conflicts, affecting subsequent retrieval efficiency. Different user inputs carry different contextual tags, such as input time, operation status, and usage path, but traditional parsing methods fail to incorporate them into a unified reference system for comparison and weighing, resulting in a fragmented semantic space. User inputs may contain similar but inconsistent keywords or phrases (e.g., "application template" vs. "download contract"), and traditional methods struggle to identify and handle potential semantic biases. Even when multiple user inputs are semantically consistent, their linguistic expressions still differ significantly, and existing technologies cannot automatically generate a set of expressions that both cover the intents of multiple parties and are suitable for structured queries. System-generated query expressions often fail to consider the current task context or system preset goals, becoming disconnected from the real business environment, causing search results to deviate from user expectations. Therefore, this paper proposes a method and system for searching by merging user input information to address the above problems. Summary of the Invention

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for performing a search by merging user input information includes the following steps:

[0008] It receives raw input information from multiple users and transforms each input into an input semantic structure through semantic parsing. This structure includes keywords, semantic relationships, and contextual labels.

[0009] Cross-compare multiple input semantic structures to identify recurring or semantically similar keywords and phrases, extract interactive focus content, and generate semantic consensus scores based on interactive focus coverage.

[0010] Based on semantic consensus scoring, and combining the keyword weights formed by each user in historical usage records with the semantic fit in the current scenario, an input fusion path graph is constructed. This path graph is used to determine the fusion order and fusion method between input semantic structures.

[0011] The fusion actions are executed sequentially along the input fusion path graph. For each pair of input semantic structures, semantic structure comparison, context consistency comparison, and semantic conflict mediation are performed. When the semantic conflict reaches an irreconcilable level, the expression is unified by using alternative words to generate a comprehensive semantic expression structure.

[0012] The keywords in the comprehensive semantic expression structure are sorted according to semantic level and task-oriented importance, a comprehensive search expression is constructed, the query operation is executed, and the search results are output.

[0013] In a preferred embodiment, when each input is transformed into the content of the input semantic structure through semantic parsing, the context labels in the input semantic structure include the following four categories: the timestamp when the input information is received, the current activity state of the user's operating system, the keyword sequence extracted from the user's three consecutive operation paths before receiving the input information, and the task function type marker in the user's current interface; the above four categories of context labels together constitute the user's current context label vector in the context of the current task.

[0014] The above four types of context labels together form the data set on which the current task background is based, which is used to constrain the parsing boundary of keywords and semantic relationships, so that the input semantic structure has a unified background reference dimension during the cross-comparison process; the timestamp corresponds to the task execution cycle node, the operation status corresponds to the functional stage, the operation path represents the causal chain of behavior, and the function type label defines the main direction of the current task intent.

[0015] In a preferred embodiment, when generating a semantic consensus score, the semantic consensus score is based on an undirected keyword graph constructed from the interactive focus content, where each node in the graph is a standardized keyword and the edges represent the frequency of co-occurrence of the keyword in different input semantic structures;

[0016] Let node weight be variable X and edge frequency be variable Y, then the semantic consensus score S is defined as: S = f(X) × P(Y), where f(X) is the normalized graph embedding vector magnitude and P(Y) is the co-occurrence edge weight defined based on Markov transition probability;

[0017] Specifically, P(Y) = ∑p(i→j), where p(i→j) is the conditional probability of continuous co-occurrence between any two keyword nodes in all semantic structures, and the consensus score of all nodes in the graph is used to select the priority path in the construction of the fusion path graph.

[0018] In a preferred embodiment, the input fusion path graph is constructed based on the set of all input semantic structures that participate in forming the semantic consensus score S. A fusion evaluation is performed on the user input corresponding to each structure in the set. The fusion evaluation result F is jointly determined by the semantic consensus score S, the keyword weight W of the user input, and the semantic fit M, where F = s × sigmoid(W + M), where s is all the keywords contained in the user input semantic structure, and the average of the semantic consensus scores obtained when there are matching nodes in the interaction focus graph; the keyword weight W represents the average of the product of the frequency of use and click consistency of all keywords in the user's current input in historical usage records; and the semantic fit M represents the average of the cosine of the angle between the semantic vectors of all keywords in the user's current input and the current task context label.

[0019] If the fusion evaluation result F is lower than the set retention threshold T, the corresponding user input information is removed from the fusion path construction, and the remaining user input information is sorted from high to low according to the F value and arranged in sequence as the node sequence in the input fusion path graph.

[0020] In a preferred embodiment, the fusion action is performed only on user input information that has passed the fusion evaluation function F with a score not lower than the retention threshold T. That is, the set of input semantic structures corresponding to the sorted valid node sequence in the input fusion path graph. During the semantic structure comparison process, a keyword structure mapping matrix is ​​constructed for each pair of input semantic structures. The matrix rows and columns represent the top k keywords in the two input semantic structures, and the matrix elements are the structural semantic similarity scores between the pair of keywords. The scores are calculated based on the product of the semantic dependency path length and the cosine of the angle between the context vectors. The cosine of the angle between the context vectors refers to the cosine of the angle between the context semantic vectors of the two keywords. The context semantic vector is the semantic expression vector of the keyword in the semantic subspace defined by the current context label vector. When the similarity score exceeds the preset matching threshold, structure merging is performed; otherwise, the semantic conflict mediation stage is entered.

[0021] In a preferred embodiment, semantic conflict resolution employs a substitution expression processing mechanism. For any two input semantic structures containing keyword pairs that are at the same semantic position but whose semantic vector distance exceeds a set distance threshold, a semantic center vector for the keyword pair is constructed. The word with the smallest distance from the center vector in the semantic neutralization word set is selected as the substitution word for uniform expression processing of the keyword pair. The processing is only performed on conflicting keyword pairs between the two input semantic structures in the current fusion operation, while other inputs retain their original structures. The semantic neutralization word set is formed into a low-density region in the semantic space after cluster analysis to ensure that the substitution process does not introduce semantic bias.

[0022] In a preferred embodiment, when generating the comprehensive semantic expression structure, the semantic level score and task-oriented importance on which the keyword ranking is based are calculated from the following sources: the semantic level score U is the product of the keyword weight W in the semantic consensus score S calculation of each keyword in the input fusion path graph and the number of times it appears in the fusion process.

[0023] Task-oriented importance V is the cosine of the semantic angle between each keyword and the task-level context label vector; the task-level context label vector is a reference standard extracted from the system task metadata.

[0024] The final keyword ranking weight Q is obtained by exponentially normalizing the sum of U and V, i.e., Q = softmax(U+V). The softmax function is W(i) = exp(U(i) + V(i)) ÷ ∑exp(U(j) + V(j)), where i and j traverse all keywords participating in the ranking. The ranking weight W determines the output order of keywords in the comprehensive semantic expression structure, ensuring that the ranking takes into account both consensus and task orientation.

[0025] In a preferred embodiment, the comprehensive search expression is constructed using a segmented splicing generation method, comprising three parts: the main semantic segment, the conflict resolution segment, and the context supplement segment.

[0026] The main semantic segment includes the top R keywords with ranking weight Q, representing the main intent content in the fused semantic structure; the conflict mediation segment includes keywords that have participated in the semantic conflict mediation process and have been replaced by unified processing, used to supplement the explanation of the semantic consistency of the final expression; the context supplement segment includes a set of keywords that are closest to the task-level context label vector in the semantic space, used to bridge the differences between the expression structure and the task environment.

[0027] The comprehensive search expression is formed by concatenating the main semantic segment, the contextual supplementary segment, and the conflict resolution segment in that order.

[0028] In a preferred embodiment, a system for performing a search by incorporating user input information includes:

[0029] The semantic parsing module is used to receive raw input information from multiple users and transform each input into an input semantic structure through semantic parsing. This structure includes keywords, semantic relationships, and context labels.

[0030] The focus extraction module is used to cross-compare multiple input semantic structures, identify recurring or semantically similar keywords and phrases, extract interactive focus content, and generate a semantic consensus score based on the interactive focus coverage.

[0031] The path construction module is used to construct an input fusion path graph based on semantic consensus scoring, combined with the keyword weights formed by each user in historical usage records and the semantic fit in the current scenario, in order to determine the fusion order and fusion method between input semantic structures.

[0032] The structure fusion module is used to sequentially perform fusion actions along the input fusion path graph. For each pair of input semantic structures, it performs semantic structure comparison, context consistency comparison, and semantic conflict mediation. When the semantic conflict reaches an irreconcilable level, it uses alternative words to unify the expression and generate a comprehensive semantic expression structure.

[0033] The expression generation module is used to sort the keywords in the comprehensive semantic expression structure according to semantic level and task-oriented importance, construct a comprehensive search expression, execute the query operation, and output the search results.

[0034] The technical effects and advantages of this invention are as follows:

[0035] This invention enables efficient fusion of multi-user input information, thereby improving the representativeness and comprehensiveness of query results. By receiving raw input information from multiple users and transforming each input into a unified semantic structure through semantic parsing, this structure includes keywords, semantic relationships, and contextual tags, allowing for unified structural modeling of input content from different users. Subsequently, the system cross-compares these structured semantic inputs, identifying recurring or semantically similar keywords and phrases, and extracting the interactive focus content as a consensus basis to generate a semantic consensus score. This mechanism ensures that the fusion process not only considers the subjectivity of individual user input but also forms more collectively representative expressions based on the semantic overlap among multiple users, making the final query operation closer to the actual needs of most users.

[0036] This invention provides a scoring-driven input fusion path determination method, effectively improving the rationality of the fusion order and the consistency of the fusion results. Based on the generated semantic consensus score, this method combines the keyword weights formed by each user's historical usage records with the semantic fit in the current scenario to construct an input fusion path graph to guide the fusion order and method. This path graph provides a global sorting and filtering basis in the fusion process, enabling sequential operations along a preset path during fusion, reducing interference from low-relevance or semantically deviating inputs. Furthermore, guided by this path graph, each pair of input semantic structures will complete comparison, mediation, and merging steps according to consistent logic during the fusion process, thereby enhancing the coordination of the fusion results in terms of contextual understanding and semantic unity.

[0037] This invention achieves precise semantic alignment between query content and user input by constructing a comprehensive semantic expression structure and generating a comprehensive search expression. After fusing the input semantic structure, the system generates a comprehensive semantic expression structure, which is the product of multiple rounds of fusion and conflict resolution, possessing semantic integrity and contextual consistency. Based on this, the system sorts the keywords in this expression structure according to semantic level and task-oriented importance, and constructs a comprehensive search expression. This expression, as the final query instruction, accurately reflects the common semantics among multiple user inputs and the most representative query intent, enabling the search engine to more effectively capture key information when executing the query, thereby outputting more accurate and comprehensive search results. Attached Figure Description

[0038] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0039] Figure 1 This is a schematic diagram of a method for searching by merging user input information according to the present invention.

[0040] Figure 2 This is a schematic diagram of a system for performing a search by merging user input information, as described in this invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Reference Figure 1 - Figure 2 The following examples were obtained:

[0043] Example 1:

[0044] A method for performing a search by merging user input information includes the following steps:

[0045] The system receives raw input from multiple users and transforms each input into a semantic structure through semantic parsing. This structure includes keywords, semantic relationships, and context labels. In natural language expression, each user's input differs in word choice, expression order, and content granularity. Processing each input through semantic parsing effectively abstracts raw, unstructured information into a semantic structure with a standard expression format. This structure contains three core components: keywords represent the information topic; semantic relationships describe the logical or causal connections between keywords; and context labels characterize the environmental state at the time of input, such as task background, user interface, and time point. This structure not only serves as the basic data unit for subsequent processing but also ensures the comparability and fusion potential of different user inputs within the same semantic space.

[0046] The system cross-compares multiple input semantic structures to identify recurring or semantically similar keywords and phrases, extracts interactive focus content, and generates a semantic consensus score based on the coverage of interactive focus content. This step aims to discover the semantic intersection between different user inputs and quantify the degree of consensus within this intersection. By cross-comparing keywords and phrases in multiple input semantic structures, the system can identify content that recurs or is semantically similar across multiple user expressions, thereby extracting a set of content representing the current multi-user input interactive focus. Interactive focus content reflects, to some extent, the core concerns or expression trends of the user group. The system further generates a semantic consensus score based on the coverage of interactive focus content across all semantic structures, quantifying the semantic strength of this consensus. This score serves as the basis for fusion judgment and ranking in subsequent steps, playing a guiding role.

[0047] Based on the semantic consensus score, and combining the keyword weights formed by each user in their historical usage records with the semantic fit in the current scenario, an input fusion path graph is constructed. This path graph is used to determine the fusion order and fusion method between input semantic structures. By jointly evaluating the "group semantic consensus strength" and the "adaptability between individual input and the current task," a fusion scoring mechanism that can be used for ranking and filtering is established. The semantic consensus score generated in the previous step has established a node embedding model based on an undirected keyword graph for each keyword. The collective semantic connectivity between keywords is jointly defined by node weights and co-occurrence edge frequencies. This score reflects the degree of consensus between input semantic structures. Building on this, the current step introduces the weight information of keywords formed by each user's input in their historical usage records to quantify the stability and usage tendency of the user's keywords. At the same time, semantic fit is introduced, representing the degree of semantic matching between keywords and the context labels of the current task, to capture the applicability of the input in the current scenario. These three factors are fused into a joint evaluation function, which serves as the basis for determining whether to participate in the subsequent fusion process and their ranking weight in the fusion path. The fusion path graph consists of all input semantic structures whose evaluation scores exceed a preset retention threshold, and is arranged in a sequence from high to low scores, serving as the node chain for the next fusion action.

[0048] The fusion process proceeds sequentially along the input fusion path graph. For each pair of input semantic structures, semantic structure comparison, contextual consistency comparison, and semantic conflict mediation are performed. When semantic conflicts reach an irreconcilable level, substitution words are used for unified expression processing, generating a comprehensive semantic expression structure. After the fusion path graph is determined, this step begins to perform fusion actions between semantic structures pair by pair. This process is not only a content-level merging but also emphasizes semantic consistency and contextual coordination. First, semantic structure comparison is performed to identify overlapping or conflicting parts of the two input structures in terms of keywords and semantic relationships. Second, the consistency of their contextual labels is evaluated to avoid background mismatch. If a semantic conflict is detected, the system attempts to mediate. When mediation fails, i.e., the semantic difference exceeds the system's set tolerance threshold, substitution words are used to select semantically neutral words to unify the expression of the keywords. Finally, a comprehensive semantic expression structure is generated. It is a unified expression after integrating all valid input semantic information, possessing semantic integrity and consistency, and is the sole source for constructing subsequent search expressions.

[0049] Keywords in the comprehensive semantic expression structure are sorted according to semantic level and task-oriented importance to construct a comprehensive search expression, execute the query operation, and output the search results. To transform the fused semantic expression into an executable search form, two dimensions for keyword sorting are defined: semantic level and task-oriented importance. The former reflects the node weight and frequency of occurrence of keywords in the consensus graph, while the latter measures the degree to which keywords fit the current task intent. The sorted keywords are used to construct the comprehensive search expression, which includes three parts: a main semantic segment, a conflict resolution segment, and a contextual supplement segment. These parts are combined in a fixed concatenation order to form a standard expression format that can be directly used in the query system. Finally, the system executes the search operation and outputs the most relevant search results to the multi-user semantic fusion results, achieving the goal of semantically unified retrieval of group intent.

[0050] In the method described in this invention, during the process of transforming each input into an input semantic structure through semantic parsing, it is necessary to extract information and semantically model the context of the input. Therefore, the context labels in the input semantic structure include the following four categories:

[0051] The first category is the "timestamp of the input information being received," which is the specific time information when the user's input was received by the system. It is usually represented by a standard date and time stamp, such as "2025-05-30 14:26:35." This timestamp not only identifies the system's receipt time of the input but can also be used to determine which stage of the task flow the input belongs to, thus corresponding one-to-one with the task execution cycle nodes. For example, in a workflow-type task, the input at the initial stage and the input at the final stage may have completely different focuses; the system can identify the temporal characteristics of the input behavior through the timestamp.

[0052] The second category is "the current active state of the user's operating system," which refers to the system's running state or interface state when the user performs input, such as editing, browsing, searching, or submitting. This state information is used to map the functional stage of a task, such as "data entry," "parameter configuration," or "result confirmation," and can help the system infer the operational context of the input intent. The operating system's active state is output by the system state management module and is usually represented by function labels, interface identifiers, or operation codes.

[0053] The third category is "keyword sequences extracted from the user's three consecutive steps before receiving input information." This refers to the keywords or objects involved in the three consecutive actions performed by the system before the user generates the current input. For example, if a user clicks "Find Customers," "View Contracts," and "Add Record" consecutively, the system will extract the keywords "customers," "contracts," and "add" from these three actions, forming a keyword sequence that describes the causal chain of the user's behavior. This sequence is behaviorally leading and can provide a semantic expectation context for the current input.

[0054] The fourth category is "task function type marker in the user's current interface," which refers to the function type label that is activated or focused on in the system interface when the user is inputting information, such as "expense approval," "report generation," and "parameter configuration." This marker semantically defines the primary task orientation of the current input intent, i.e., what system task goal the input information serves. This marker originates from the system task scheduling module or the interface interaction management module.

[0055] The four types of context labels mentioned above together form the current context label vector of the user in the context of the current task. This vector is a high-dimensional semantic vector, representing the semantic environment state in which the user inputs. Each label dimension is converted into a vector through a pre-trained semantic embedding model (such as BERT, Word2Vec, etc.) and then merged to form a unified input for semantic calculation and comparison.

[0056] Furthermore, the four types of context labels mentioned above together constitute the dataset upon which the current task context is based, constraining the boundaries of keyword and semantic relationship recognition in subsequent semantic parsing processes. Specifically, when performing parsing operations such as keyword extraction, relationship determination, and semantic path generation, operations are only performed within the semantic subspace defined by the current context label vector. This ensures that the input semantic structures generated from different user inputs have a unified background reference dimension, avoiding discrepancies in parsing structures due to differences in contextual environments.

[0057] Taking an e-commerce platform as an example, if user A clicks on "Add Product," "Set Price," and "Set Inventory" in the "Product Management" interface at 10:00 AM on May 30th, and then enters the "Parameter Configuration" interface and inputs the text "Set as Default for Discount Season," the context tags entered by the user are as follows: Timestamp: 2025-05-30 10:00:23; Current Activity Status: Function Module marked as "Parameter Configuration in Progress"; Operation Path Keyword Sequence: [Add, Price, Inventory]; Current Function Type Marker: Discount Settings Module. Based on this, the system generates a context tag vector that focuses on keywords such as "product operation" and "discount configuration," enabling the semantic parsing module to more accurately understand that the word "default" refers to the discount template, rather than other parameter items.

[0058] It should be noted that the current context label vector does not directly extract keywords to participate in constructing the semantic structure. Instead, it serves as a semantic parsing boundary control factor, limiting the semantic space scope of the parsing process and playing the roles of "limiting the semantic space," "filtering candidates," and "ensuring consistency." Its specific functions are as follows:

[0059] Keyword extraction stage: The current context label vector defines the semantic space of the parsed candidate words. The system only performs keyword candidate filtering on terms whose semantic similarity to the vector exceeds a preset value, so as to exclude redundant semantic units that are irrelevant to the current task intent.

[0060] Relationship determination stage: The current context label vector is used to limit the set of candidate semantic relationships. For example, in "approval" tasks, typical semantic relationship templates such as "authorization-rejection" and "initiation-processing" will be used first, so as to avoid the relationship determination process from entering a semantic paradigm that is irrelevant to the current task.

[0061] Semantic path generation stage: The search algorithm for constructing semantic paths (such as graph traversal or dependency tree expansion) only allows semantic jumps between path nodes to occur within the semantic subspace defined by the context labels, ensuring path continuity and background consistency. In summary, the current context label vector is not equivalent to "a set of keywords," but rather a high-dimensional semantic embedding vector. Its function is to limit the spatial dimension and constraint range of semantic parsing, ensuring that the semantic structure of multi-user input maintains a unified parsing standard during the generation stage, thus fundamentally guaranteeing the feasibility and consistency of cross-comparison and fusion.

[0062] In the method described in this invention, a semantic consensus score is used to measure the strength of semantic commonality among different user inputs. This score serves as a core basis for constructing the input fusion path graph, determining the fusion priority between different input semantic structures. This score is generated before the input fusion stage and is based on an undirected keyword graph constructed from interactive focus content. Specifically, the system first extracts interactive focus content based on keywords or semantically similar phrases that intersect in multiple input semantic structures. Interactive focus content refers to expression units that repeatedly appear in two or more input semantic structures or are highly semantically similar (e.g., achieving a similarity of 0.8 or higher), possessing group consensus characteristics and representing the core information that users are concerned about in the current scenario.

[0063] Based on these interactive focus contents, the system constructs an undirected keyword graph. Each node in the graph represents a standardized keyword. Standardization includes unifying synonyms, eliminating word form variations (such as plural / past tense), and merging domain terms. The edge between any two nodes in the graph represents the frequency of co-occurrence of that pair of keywords across different input semantic structures; that is, how many different user inputs contain both words simultaneously.

[0064] Based on this, a node weight X is assigned to each node. The node weight X represents the importance of the keyword's position in the interactive focus graph, which can be obtained by generating a vector using graph embedding algorithms (such as DeepWalk and Node2Vec) and then normalizing its magnitude. f(X) represents the normalized vector magnitude of the node, indicating its structural stability. Simultaneously, the edge frequency Y represents the co-occurrence strength connecting the keyword with other keywords, i.e., the frequency of the keyword appearing alongside other words in the graph during multi-user input, used to construct the co-occurrence edge weights. The system uses a Markov transition probability-based weighted modeling of the edge frequency, defined as: P(Y) = ∑p(i→j), where p(i→j) is the conditional probability that any two keyword nodes i and j appear in a fixed order (e.g., "A then B") across all semantic structures, calculated as follows:

[0065] p(i→j) = count(i followed by j) ÷ count(i appears); that is, the proportion of the frequency of keyword i followed by keyword j in all user inputs to the total frequency of i. If two words appear multiple times in consecutive positions in the semantic structure, it indicates that their semantic dependency is high and the transition probability is large. After obtaining f(X) and P(Y), the semantic consensus score S is calculated, which is defined as follows: S = f(X) × P(Y), that is, the consensus score S is determined by the product of the structural stability of a keyword node in the embedding space of the graph structure and its transition strength in the semantic co-occurrence network. The higher the S value, the stronger the consensus of the keyword in multi-user input. Finally, the system uses the consensus score results of all nodes in the graph for priority path selection in the construction of the input fusion path graph. When constructing the path graph, the system will refer to whether the keywords contained in each input semantic structure match the high-scoring consensus keyword nodes, and adjust the arrangement order of nodes in the fusion path accordingly.

[0066] Taking the configuration of a product recommendation system by multiple users as an example, user A inputs "set recommendation logic", user B inputs "configure content push", and user C inputs "update push rules". After parsing, the system can extract keywords such as "set", "configure", "recommend", "push", and "rules". It finds that "push" appears in all three users or has similar semantics (such as "content push", "recommendation push", "push rules"). Based on this, the system constructs an interaction focus graph. If "push" and "recommend" co-occur continuously in users A and B, the system sets the frequency of that edge to a higher value and calculates p(recommendation→push) as 0.75 and p(push→recommendation) as 0.32. Taking the "push" node as an example, the system obtains its vector magnitude from the graph embedding model as 0.68. Then, the consensus score of the keyword "push" is: S = f(X) × P(Y) = 0.68 × (0.75 + 0.32) = 0.73 (approximate). This score will be used as the basis for ranking and fusion judgment in the subsequent input fusion path construction.

[0067] In the method described in this invention, the construction of the input fusion path graph is based on the semantic consensus score generated in the previous stage. It further integrates the historical behavioral features of each user input with the current task semantic environment to determine and rank the fusion value of the user input information. The starting point of the construction process is the set of all input semantic structures that participate in forming the semantic consensus score S. This set contains all user input information that the system has received and completed semantic parsing processing. Each user input is transformed into an input semantic structure through semantic parsing, containing three core elements: keywords, semantic relationships, and context labels. These structures are uniformly embedded into the same semantic vector space to ensure a consistent basis for subsequent comparisons and fusion operations. For each input semantic structure in the above set, the system performs a fusion evaluation operation to quantify its contribution to the final comprehensive expression structure construction. The fusion evaluation result F is the quantitative output of this evaluation, defined as follows: F = s × sigmoid(W + M); where: variable s refers to the subset of keywords contained in the user input semantic structure that have matching nodes in the interaction focus graph constructed in the previous stage, and the value obtained by averaging the semantic consensus score values ​​corresponding to this subset. This value reflects the degree of alignment between the user's input and the semantic focus of group interaction, and has a global consensus orientation.

[0068] Keyword weight W is defined as the average of the product of the frequency of use and click consistency of all keywords in the user's current input in historical usage records. Frequency of use represents how often a keyword appears in the user's past interactions, while click consistency represents the degree of consistency between the actions triggered by the keyword (such as clicking, confirming, browsing, etc.) and the system's expected task path. The product of these two factors constitutes a measure of the "behavioral value" of the keyword.

[0069] Semantic fit M is defined as the average of the cosines of the angles between all keywords input by the user and the current task context labels in the semantic vector space. This value indicates whether the input content fits the current task context; if the semantic vector of a keyword is similar in direction to the context label vector, its contribution is higher.

[0070] In actual calculations, the system first extracts the vector representation of each keyword through semantic parsing, then extracts its corresponding historical records from the user behavior database, and finally calculates the W and M values ​​respectively. The sigmoid function is then used to perform non-linear normalization to ensure that the output score range is controlled and has discriminative power. Once the fusion evaluation value F is calculated, the system compares this value with a preset retention threshold T: if F is lower than the threshold T, it indicates that the input is insufficient in terms of consensus, historical validity, and task adaptability, and the system will remove this input information from the input fusion path graph construction; if F is higher than or equal to the threshold T, the input is retained, and all retained inputs are sorted from highest to lowest according to their F values, arranged sequentially as a node sequence in the input fusion path graph. This node sequence in the path graph not only represents the order in which each user input structure participates in the fusion, but also determines the pairing logic and mediation process of the fusion action in the subsequent structural fusion stage. Through the above fusion evaluation mechanism, this invention effectively filters redundant and off-target user inputs, retaining only high-semantic-quality and representative input data, enhancing the stability and accuracy of the comprehensive semantic expression structure.

[0071] Taking a real-world scenario as an example: User A inputs "Set recommendation rules", User B inputs "Optimize push configuration", and User C inputs "Update marketing logic". The system performs semantic parsing on the three inputs, extracting keywords such as "settings", "recommendation", "push", "configuration", and "marketing" to form their respective semantic structures. The interaction focus graph shows that "recommendation", "push", and "configuration" frequently co-occur in multiple inputs, thus the corresponding keywords obtain a high consensus score S in the graph. Assuming that in User B's input, "push" and "configuration" both match the nodes in the interaction focus graph, the resulting S is 0.85; at the same time, this user has a high historical usage frequency of the keyword "push" and often uses it in the push function module, resulting in a high W value, set to 0.73; the semantic similarity between the keywords in their input and the current context label (such as "content operation interface") is high, with an M value of 0.68; then:

[0072] F = 0.85 × sigmoid(0.73 + 0.68) ≈ 0.85 × sigmoid(1.41) ≈ 0.85 × 0.80 ≈ 0.68. If the system-set retention threshold T is 0.65, then user B's input is retained and obtains a higher position in the input fusion path graph. A similar evaluation method is applicable to all user inputs, forming a fusion path graph with stable ranking, clear source, and semantic coherence.

[0073] In the method described in this invention, the fusion action is only performed on user input information that has passed the fusion evaluation function F with a score not lower than the retention threshold T. As mentioned above, the fusion evaluation function F reflects the comprehensive semantic value of a certain input semantic structure, and it can only enter the subsequent fusion processing stage when its F value exceeds the preset threshold T. All input semantic structures that meet this condition are sorted from high to low according to their F values, forming a valid node sequence in the input fusion path graph.

[0074] The set of input semantic structures corresponding to the valid node sequence in the input fusion path graph is the set of target objects for subsequent structural fusion operations. The system will sequentially extract adjacent structural pairs according to the arrangement order of the structures in the sequence, and perform semantic structure comparison one by one to determine whether they have sufficient semantic structure similarity, thereby deciding whether to merge or enter the conflict mediation process.

[0075] During semantic structure comparison, the system constructs a keyword structure mapping matrix for each pair of input semantic structures. This matrix is ​​a two-dimensional mapping table, where the rows and columns represent the top k (e.g., 10) keywords in the two input semantic structures. Each element in the matrix represents the structural-level semantic similarity score between the pair of keywords, which quantifies the degree of matching between keywords in two different structures in terms of semantic logical structure. Specifically, the formula for calculating the structural-level semantic similarity score is: Similarity Score = Semantic Dependency Path Length × Cosine of the Angle Between Context Vectors; Semantic Dependency Path Length: refers to the length of the semantic dependency path between the keyword to be compared and other keywords in each semantic structure. The shorter the path, the stronger the semantic core of the word in that structure. Cosine of the Angle Between Context Vectors: refers to the cosine of the angle between the context semantic vectors of the two keywords, used to measure whether their semantic environments are similar. The closer the value is to 1, the more consistent the semantic environments.

[0076] By multiplying these two metrics, the resulting score integrates semantic status and contextual consistency, and can be used to assess whether the keyword pair has structural equivalence in two semantic structures. For all scoring items in the matrix, if any score exceeds the preset matching threshold set by the system, the system considers the pair of input semantic structures to have the feasibility of structural fusion.

[0077] Once the above criteria are met, the system will perform a structure merging operation, which involves unifying identical or similar keyword nodes in two semantic structures and merging, simplifying, and uniformly encoding their associated semantic relationships and contextual tags. The purpose of the structure merging operation is to consolidate repetitive information in the semantic expression space, thereby improving the conciseness and consistency of the overall structure.

[0078] If a pair of input semantic structures does not contain a keyword pair with a score exceeding the matching threshold, it indicates a significant difference in their semantic structures. In this case, the system will not force structural fusion but will instead guide the pair into the semantic conflict resolution stage. The conflict resolution mechanism will be executed in subsequent processes, such as by using alternative expressions or generating semantically neutralizing words.

[0079] For example, consider two adjacent structures in the input fusion path graph. Assume user A's input is "optimize product recommendation logic," and user B's input is "configure recommendation content parameters." After semantic parsing, the first k keywords are: Structure A: optimization, product, recommendation, logic; Structure B: configuration, recommendation, content, parameters. The system constructs a keyword structure mapping matrix, with rows and columns corresponding to the first four keywords in each structure. Pairwise comparisons reveal that the keyword pair "recommendation—recommendation" has a context semantic vector cosine value of 0.94, semantic dependency paths of 1, and a product score of 0.94, exceeding the preset matching threshold of 0.80. Therefore, this keyword pair is determined to be a structurally fusionable pair, and its corresponding semantic relationships, such as "subject-verb" and "object," will be merged. Other keywords, such as "optimization—configuration" and "logic—parameter," lack sufficient similarity and do not meet the fusion criteria, requiring further adjustment.

[0080] It should be noted that the context semantic vector is a context constraint vector representation derived by combining the current context label vector with the keyword semantic vector in the modeling process. It can be viewed as the semantic expression vector of the keyword within the semantic subspace defined by the current context label vector. The current context label vector consists of four types of context labels: timestamp, operating system state, keyword sequence of user operation path, and task function type marker, representing the semantic boundary of the current task background. This vector participates in multiple stages such as semantic parsing and semantic fit calculation. The context semantic vector measures the specific semantic position of a keyword in the current semantic environment; it is essentially derived by jointly modeling the keyword's basic semantic vector and the current context label vector, representing the keyword's expression result in the context space. Contextual semantic vectors refer to the semantic embedding representation generated by a keyword in its semantic structure and combined with the current context label vector. They are used to reflect the actual semantic projection result of the keyword in the current task scenario. The current task scenario refers to the semantic environment context that the system records and extracts when it receives each piece of user input information, which is directly related to the input behavior. Specifically, it includes feature factors such as the user's functional interface, task stage, behavior path, and interaction time, which determine the goal or task intent to be achieved by the input.

[0081] In the method of this invention, to ensure the consistency and stability of the comprehensive semantic expression structure generated during the fusion process of different user input information, a semantic conflict mediation mechanism is specifically designed. This mechanism is applicable to situations where, during the structural fusion process, two input semantic structures cannot be directly merged due to differences in semantic expression. This mediation mechanism adopts an alternative expression processing mechanism, aiming to reconcile the differing words in the two semantic structures by finding suitable neutral expression terms in the semantic space, thereby achieving semantic unity and avoiding expression bias caused by direct discarding or forced merging. The judgment criteria for semantic conflict are as follows: When the system performs the fusion action, it analyzes the keywords inside each of the two adjacent input semantic structures in the current input fusion path graph. If it detects that the distance between the semantic vectors of two keywords in the same semantic position in the high-dimensional semantic space exceeds a set distance threshold, then the keyword pair is determined to have a semantic conflict.

[0082] Keyword pair: refers to a set of keywords in two input semantic structures that correspond in terms of grammatical position, semantic role, etc., such as "customer" and "user" in the subject position; Same semantic position: refers to two keywords having similar dependency grammatical roles or semantic labels in their respective structures, such as subject to subject, verb to verb, etc.; Semantic vector distance: defined by the Euclidean distance or cosine angle between vectors generated by the embedding model (such as Word2Vec, BERT, etc.); Set distance threshold: the maximum acceptable semantic deviation value predefined by the system, used to determine whether to trigger the mediation mechanism. Once a semantic conflict between keyword pairs is identified, the system will perform the following mediation steps:

[0083] Constructing the semantic center vector: The semantic vectors of conflicting keyword pairs are averaged or weighted (e.g., weighted based on the semantic level score of the keywords) to obtain the semantic center vector. This vector lies between the two words and represents a potential neutral semantic direction.

[0084] Retrieving candidate terms from a semantically neutral term set: The system pre-forms a semantically neutral term set through cluster analysis in the semantic space. This set consists of a series of words with relatively ambiguous semantic expressions, low semantic density, and broad applicability across multiple tasks. This set often includes general terms such as "object," "participant," and "target item."

[0085] Selecting the term with the smallest distance as the replacement word: From the semantic neutral word set, the system calculates the distance between the semantic vector of each term and the semantic center vector in step 1, and selects the term with the smallest distance as the replacement word for the keyword pair, thereby achieving a unified semantic expression without obviously favoring either original expression.

[0086] Replace conflicting keywords in the original structure: The system only replaces the keyword pair in the two input semantic structures involved in the current fusion operation. Other input semantic structures that are not involved in the current fusion operation will retain their original structure, ensuring that the fusion behavior is local and does not spread to non-conflicting areas.

[0087] Continue with the structural fusion: After unifying the alternative words, rebuild the keyword mapping matrix of the structural pair. If there are no conflicts at this point, the structural merging is completed; otherwise, proceed to the next round of mediation.

[0088] In the search method proposed in this invention, which combines user input information, the keyword ranking mechanism plays a crucial role in the generation of the comprehensive semantic expression structure. To ensure that the ranking results achieve a balance between expression consistency and task orientation, the system ranks keywords based on two core scoring indicators: semantic level score U and task orientation importance V.

[0089] The composition and calculation of the semantic ranking score U: The semantic ranking score U of a keyword reflects the degree of semantic consensus it embodies in the user group's input, derived from its overall contribution to the semantic consensus score calculation. Specifically: U = cumulative keyword weight W × frequency of occurrence; where: the cumulative keyword weight W refers to the direct summation of the keyword weight W values ​​assigned to each input when a specific keyword appears in all user inputs participating in the construction of the semantic consensus score S; the keyword weight W itself refers to the average of the product of the frequency of use and click consistency of all keywords in the user's current input in the historical usage record; since multiple user inputs may contain the same keyword simultaneously, the keyword will obtain multiple W values, which are accumulated and used as the first item in the semantic ranking score U; the frequency of occurrence refers to the cumulative frequency of occurrence of the keyword during the fusion process when it is structure-matched, extracted, involved in conflict mediation, or structure-ranked. For example, the keyword "configuration" was identified as a high-weight keyword in the inputs of 5 users, with corresponding keyword weights W values ​​of 0.81, 0.77, 0.92, 0.68, and 0.88, respectively. The cumulative W value for the keyword "configuration" is: W_total = 0.81 + 0.77 + 0.92 + 0.68 + 0.88 = 4.06. If this keyword appears a total of 7 times in the structural fusion stage, its semantic level score is: U = 4.06 × 7 = 28.42. This score will be used as the first dimension score in the subsequent ranking.

[0090] The composition and calculation of task-oriented importance V: The task-oriented importance V of a keyword is used to measure the degree of fit between its semantic orientation and the current task goal of the system. The calculation method is as follows: V = cosine value of the semantic angle between the keyword semantic vector and the task-level context label vector, where: the task-level context label vector is a semantic embedding vector generated from a set of label content (such as task goal, functional module category, system business node, etc.) extracted from the system task metadata; the angle calculated with it is the embedding vector of each keyword to be ranked in the semantic space; the closer the cosine similarity value is to 1, the better the keyword reflects the current task goal, and the higher the ranking priority.

[0091] It should be noted that the two-level structure of contextual tags is as follows: In order to take into account both the user input context and the system task background, this invention introduces a two-level structure of contextual tags:

[0092] User-level context labels: Labels attached to each user input; including input timestamp, operation path keyword sequence, current system status, and task function type marker; used to constrain the boundaries and scope of the user input during the semantic parsing phase;

[0093] Task-level context labels: extracted from the system's global task metadata; used to unify the sorting direction of keywords in the comprehensive semantic expression structure; and involved in calculating the angle between keywords and the semantic alignment of system tasks.

[0094] The final keyword ranking weight Q is calculated by summing the two scoring dimensions mentioned above and then processing the result using the softmax function: Q = softmax(U + V). The calculation iterates through all keywords involved in the ranking. The softmax function ensures that the keyword ranking weights are uniform, giving each keyword a clear relative ranking in the overall semantic expression structure. This ranking mechanism retains the consistency (U) gained from user input and strengthens the alignment (V) with the system's task objectives, enabling the generated search expression to more accurately guide downstream query behavior and avoid semantic dispersion and task deviation.

[0095] In the method for searching by merging user input information proposed in this invention, the comprehensive search expression is constructed using a segmented concatenation generation method to ensure the integrity, consistency, and task orientation of the merged semantic information. The expression consists of three parts: a main semantic segment, a conflict resolution segment, and a contextual supplement segment, which are concatenated in this order to construct an expression result with precise query capabilities.

[0096] The main semantic segment represents the set of keywords with the highest consensus and task orientation in the fused input semantic structure, and is the core component of the comprehensive search expression. Source: Keywords ranked first R by their ranking weight Q are extracted from the fused semantic structure; the keywords in this segment generally express the main entities, actions, or constraints of the search intent; Example: If the user input focuses on "contract template application," "authorization signing," and "auto-submission," the top R items with ranking weights might be "authorization," "template," and "auto-submission."

[0097] The conflict resolution section is used to supplement and explain semantic conflicts that arose during the initial fusion process but have been uniformly processed, thereby improving semantic consistency and system interpretability. Source: All keywords involved in the semantic conflict resolution process; Keyword selection criteria: Terms identified as conflicting keyword pairs during the fusion operation and uniformly expressed through a substitution mechanism; Substitution processing method: Constructing a semantic center vector for the conflicting keyword pair, and selecting the term with the smallest distance from the semantic neutralization word set as the final substitution expression; This section is not used for core retrieval control but contributes to the semantic integrity of the search expression; Example: If multiple users use "contract template download" and "template application form" respectively, and during structural fusion, "application form" and "download" are replaced with the neutralization word "obtain," then "obtain" will be included in this section.

[0098] The context supplement segment enhances the semantic alignment between the search expression and the task environment, ensuring that the query results align with the current task objective. Source: The set of keywords with the shortest semantic distance to the task-level context label vector; the task-level context label vector is extracted from system task metadata and serves as a global reference background for the current query intent; the keywords in this segment may not be explicitly present in the user input, but are generated from the system task structure or typical corpus through contextual reasoning mechanisms; Example: If the current task is "contract archiving," but the user has not explicitly mentioned the keyword "archiving," the system will supplement the context label vector with terms such as "archiving," "process," and "categorization" to increase the task coverage of the search intent.

[0099] The final comprehensive search expression follows this assembly order: main semantic segment → context supplement segment → conflict resolution segment. This order is designed to: prioritize task centrality of the expression structure; fill in expression gaps in contextual understanding; and finally, integrate conflicting content after unifying the expression to ensure semantic consistency. Furthermore, each segment is automatically generated based on structured semantic elements, and the expression supports standard query interfaces, such as structured retrieval requests based on Boolean expressions or natural language conversion. This mechanism enhances the accuracy of search results and the matching degree of user intent, while also improving the system's responsiveness to multi-user input fusion queries.

[0100] Example 2: A system for performing a search by merging user input information, comprising:

[0101] The semantic parsing module is used to receive raw input information from multiple users and transform each input into an input semantic structure through semantic parsing. This structure includes keywords, semantic relationships, and context labels.

[0102] The focus extraction module is used to cross-compare multiple input semantic structures, identify recurring or semantically similar keywords and phrases, extract interactive focus content, and generate a semantic consensus score based on the interactive focus coverage.

[0103] The path construction module is used to construct an input fusion path graph based on semantic consensus scoring, combined with the keyword weights formed by each user in historical usage records and the semantic fit in the current scenario, in order to determine the fusion order and fusion method between input semantic structures.

[0104] The structure fusion module is used to sequentially perform fusion actions along the input fusion path graph. For each pair of input semantic structures, it performs semantic structure comparison, context consistency comparison, and semantic conflict mediation. When the semantic conflict reaches an irreconcilable level, it uses alternative words to unify the expression and generate a comprehensive semantic expression structure.

[0105] The expression generation module is used to sort the keywords in the comprehensive semantic expression structure according to semantic level and task-oriented importance, construct a comprehensive search expression, execute the query operation, and output the search results.

[0106] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0107] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0108] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0109] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0110] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for searching by merging user input information, characterized in that, Includes the following steps: It receives raw input information from multiple users and transforms each input into an input semantic structure through semantic parsing. This structure includes keywords, semantic relationships, and contextual labels. Cross-compare multiple input semantic structures to identify recurring or semantically similar keywords and phrases, extract interactive focus content, and generate semantic consensus scores based on interactive focus coverage. Based on semantic consensus scoring, and combining the keyword weights formed by each user in historical usage records with the semantic fit in the current scenario, an input fusion path graph is constructed. This path graph is used to determine the fusion order and fusion method between input semantic structures. The fusion actions are executed sequentially along the input fusion path graph. For each pair of input semantic structures, semantic structure comparison, context consistency comparison, and semantic conflict mediation are performed. When the semantic conflict reaches an irreconcilable level, the expression is unified by using alternative words to generate a comprehensive semantic expression structure. Keywords in the comprehensive semantic expression structure are sorted according to semantic level score and task-oriented importance. The semantic level score is the product of the cumulative keyword weight in the semantic consensus score calculation and the number of times it appears during the fusion process. The task-oriented importance is the cosine of the semantic angle between each keyword and the task-level context label vector, which is a reference standard extracted from the system task metadata. A comprehensive search expression is constructed, the query operation is executed, and the search results are output.

2. The method for searching by merging user input information according to claim 1, characterized in that, When each input is transformed into the content of the input semantic structure through semantic parsing, the context labels in the input semantic structure include the following four categories: the timestamp when the input information is received, the current activity state of the user's operating system, the keyword sequence extracted from the user's three consecutive operation paths before receiving the input information, and the task function type marker in the user's current interface; the above four categories of context labels together form the current context label vector of the user in the context of the current task.

3. The method for searching by merging user input information according to claim 2, characterized in that, When generating semantic consensus scores, the semantic consensus scores are based on an undirected keyword graph constructed from the interactive focus content. Each node in the graph is a standardized keyword, and the edges represent the frequency of co-occurrence of the keyword in different input semantic structures. Let node weights be variables X and edge frequencies be variables Y. Then the semantic consensus score S is defined as: S = f(X) × P(Y), where f(X) is the normalized graph embedding vector magnitude and P(Y) is the co-occurrence edge weight defined based on Markov transition probability.

4. The method for searching by merging user input information according to claim 3, characterized in that, The construction of the input fusion path graph is based on the set of all input semantic structures that participate in forming the semantic consensus score S. A fusion evaluation is performed on the user input corresponding to each structure in the set. The fusion evaluation result F is jointly determined by the semantic consensus score S, the keyword weight W of the user input, and the semantic fit M, where F = s × sigmoid(W + M), where s is the total number of keywords contained in the user input semantic structure, and the average of the semantic consensus scores obtained when there are matching nodes in the interaction focus graph. The keyword weight W represents the average of the product of the frequency of use of all keywords in the user's current input and click consistency in historical usage records. The semantic fit M represents the average of the cosine of the angle between the semantic vectors of all keywords in the user's current input and the current task context label. Click consistency refers to the degree of consistency between the behavior triggered after the word appears, including clicking, confirming, and browsing behaviors, and the system's expected task path. If the fusion evaluation result F is lower than the set retention threshold T, the corresponding user input information is removed from the fusion path construction, and the remaining user input information is sorted from high to low according to the F value and arranged in sequence as the node sequence in the input fusion path graph.

5. The method for searching by merging user input information according to claim 4, characterized in that, The fusion action only operates on user input information that has passed the fusion evaluation function F with a score not lower than the retention threshold T. That is, the set of input semantic structures corresponding to the sorted valid node sequence in the input fusion path graph. During the semantic structure comparison process, a keyword structure mapping matrix is ​​constructed for each pair of input semantic structures. The matrix rows and columns represent the top k keywords in the two input semantic structures, and the matrix elements are the structural semantic similarity scores between the pair of keywords. This score is calculated based on the product of the semantic dependency path length and the cosine of the angle between the context vectors. The cosine of the angle between the context vectors refers to the cosine of the angle between the context semantic vectors of the two keywords. The context semantic vector is the semantic expression vector of the keyword in the semantic subspace defined by the current context label vector. When the similarity score exceeds the preset matching threshold, structure merging is performed; otherwise, the semantic conflict mediation stage is entered.

6. The method for searching by merging user input information according to claim 5, characterized in that, Semantic conflict resolution employs a substitution expression processing mechanism. For any two input semantic structures containing keyword pairs that occupy the same semantic position but whose semantic vector distance exceeds a set distance threshold, a semantic center vector for the keyword pair is constructed. The word with the smallest distance to this center vector is selected from the semantic neutralization word set as a substitution word for uniform expression processing of the keyword pair. The processing is only performed on conflicting keyword pairs between the two input semantic structures in the current fusion operation; other inputs retain their original structures. The semantic neutralization word set is a collection of words with relatively ambiguous semantic expression, low semantic density, and broad applicability across multiple tasks, pre-formed through cluster analysis in the semantic space.

7. The method for searching by merging user input information according to claim 6, characterized in that, When generating the comprehensive semantic expression structure, the semantic level score and task-oriented importance on which the keyword ranking is based are calculated from the following sources: the semantic level score U is the cumulative value of the keyword weight W of each keyword in the input fusion path graph participating in the semantic consensus score S calculation, and then the product of the cumulative value and the number of times it appears in the fusion process is obtained. Task-oriented importance V is the cosine of the semantic angle between each keyword and the task-level context label vector; The task-level context label vector is a reference standard extracted from the system task metadata; The final keyword ranking weight Q is obtained by exponentially normalizing the sum of U and V, i.e., Q = softmax(U + V).

8. The method for searching by merging user input information according to claim 7, characterized in that, The comprehensive search expression is constructed using a segmented splicing generation method, consisting of three parts: the main semantic segment, the conflict resolution segment, and the context supplement segment. The main semantic segment includes the top R keywords with ranking weight Q, representing the main intent content in the fused semantic structure; the conflict mediation segment includes keywords that have participated in the semantic conflict mediation process and have been replaced by unified processing, used to supplement the explanation of the semantic consistency of the final expression; the context supplement segment includes a set of keywords that are closest to the task-level context label vector in the semantic space, used to bridge the differences between the expression structure and the task environment. The comprehensive search expression is formed by concatenating the main semantic segment, the contextual supplementary segment, and the conflict resolution segment in that order.

9. A system for performing a search by merging user input information, used to implement the method for performing a search by merging user input information as described in any one of claims 1-8, characterized in that, include: The semantic parsing module is used to receive raw input information from multiple users and transform each input into an input semantic structure through semantic parsing. This structure includes keywords, semantic relationships, and context labels. The focus extraction module is used to cross-compare multiple input semantic structures, identify recurring or semantically similar keywords and phrases, extract interactive focus content, and generate a semantic consensus score based on the interactive focus coverage. The path construction module is used to construct an input fusion path graph based on semantic consensus scoring, combined with the keyword weights formed by each user in historical usage records and the semantic fit in the current scenario, in order to determine the fusion order and fusion method between input semantic structures. The structure fusion module is used to sequentially perform fusion actions along the input fusion path graph. For each pair of input semantic structures, it performs semantic structure comparison, context consistency comparison, and semantic conflict mediation. When the semantic conflict reaches an irreconcilable level, it uses alternative words to unify the expression and generate a comprehensive semantic expression structure. The expression generation module is used to sort the keywords in the comprehensive semantic expression structure according to their semantic level scores and task-oriented importance, construct a comprehensive search expression, execute the query operation, and output the search results.

Images