A keyword search intention analysis method and system based on GEO

Abstract

The application discloses a kind of keyword search intention analysis method and system based on GEO, it is related to search engine optimization technical field, including: initial clustering structure is constructed;Based on the local sparse distribution characteristics of semantic matching relationship between initial clustering structure node and keyword, extract local abnormal node area of semantic matching relationship;According to the topological structure asymmetric characteristics of context path between nodes in local abnormal node area, reconstruct the semantic path between nodes, form optimized candidate semantic path set;In combination with the non-stationary characteristics of state transition topological structure of context sequence in candidate semantic path set, determine the subdivision search intention corresponding to target keyword by the path state inference of conditional random field model;The application realizes the analysis of the multi-context semantic structure of search keyword, and is conducive to the subdivision identification of search intention under generative engine optimization environment.

Description

Technical Field

[0001] This invention relates to the field of search engine optimization technology, specifically to a keyword search intent analysis method and system based on GEO. Background Technology

[0002] With the continuous development of search engine technology, accurately analyzing and understanding users' search intent has become a key aspect of improving information retrieval quality. Traditional keyword search analysis methods largely rely on static semantic matching, making it difficult to effectively identify the rich and diverse intent categories behind keywords. Especially in the context of Generative Engine Optimization (GEO), the semantic complexity of search keywords has increased significantly, making it difficult for traditional semantic analysis models to fully uncover the implicit and deep relationships between keywords and users' search purposes, thus affecting the accuracy of search results and user experience.

[0003] Existing methods for analyzing search intent typically identify intent through fixed-pattern keyword clustering or simple semantic mapping, neglecting the contextual features and non-linear semantic path relationships in historical keyword usage. Therefore, they struggle to effectively address the ambiguity and vagueness of keyword search intent in real-world scenarios, particularly failing to fully utilize the dynamic and heterogeneous semantic information embodied in historical keyword search data, thus limiting search engines' ability to finely identify users' potential intent.

[0004] Therefore, in the Generative Engine Optimization (GEO) environment, there is an urgent need to explore a new method that can fully mine and integrate the local anomaly distribution features and contextual semantic connection patterns in the historical semantic data of keywords, so as to further improve the accuracy of search keyword intent analysis. Summary of the Invention

[0005] The purpose of this invention is to provide a keyword search intent analysis method and system based on GEO to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a keyword search intent analysis method based on GEO, characterized in that it includes: Semantic fragments are analyzed based on historical data of the target search keywords, and an initial clustering structure is constructed by analyzing the local deviation features of node density in the contextual co-occurrence patterns between semantic fragments. Based on the local sparse distribution features of semantic matching relationships between nodes in the initial cluster structure and keywords, local abnormal node regions of semantic matching relationships are extracted. Based on the asymmetric topological characteristics of the contextual paths between nodes within the local abnormal node region, the semantic paths between nodes are reconstructed to form an optimized set of candidate semantic paths. By combining the non-stationary characteristics of the state transition topology of the context sequence in the candidate semantic path set, the subdivided search intent corresponding to the target keyword is determined through path state inference using a conditional random field model.

[0007] Secondly, the present invention provides a keyword search intent analysis system based on GEO, implemented based on the method described above, including: The building module is used to parse semantic fragments based on historical data of target search keywords and construct an initial clustering structure through the local deviation features of node density in the contextual co-occurrence patterns between semantic fragments. The extraction module is used to extract the local sparse distribution features of the semantic matching relationship between the nodes of the initial cluster structure and the keywords, and to extract the local abnormal node regions of the semantic matching relationship. The optimization module is used to reconstruct the semantic paths between nodes based on the asymmetric topological features of the context paths between nodes within the local abnormal node region, thereby forming an optimized set of candidate semantic paths. The determination module is used to combine the non-stationary features of the state transition topology of the context sequence in the candidate semantic path set, and determine the subdivided search intent corresponding to the target keyword through path state inference by the conditional random field model.

[0008] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention reconstructs the semantic structure of historical search data by introducing node density local deviation features based on contextual co-occurrence patterns. This allows the association between basic semantic units to no longer be limited to uniformly distributed semantic connections, thereby characterizing the semantic clustering differences of keywords in different contexts and providing a more discriminative structural basis for subsequent segmentation of search intent.

[0009] This invention combines the local sparse distribution features of semantic matching relationships to screen and identify basic semantic units in the initial clustering structure, effectively revealing the inconsistency between semantic matching strength and structural connection features. This enables the extraction of key node regions that deviate from the relevance of the target search keywords in complex semantic environments, enhancing the ability to identify multi-intent cross-scenes.

[0010] This invention reconstructs semantic paths based on asymmetric features of context path topology and combines a path state inference mechanism based on state transition boundaries to perform serialization analysis on candidate semantic paths. This allows the intent transition relationships in different semantic paths to be fully expressed, thereby enabling the determination of search intent for subdivided search keywords in a generative engine optimization environment. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0012] Figure 1 This is a flowchart of a keyword search intent analysis method based on GEO according to the present invention; Figure 2 This is a framework diagram of a keyword search intent analysis system based on GEO according to the present invention. Detailed Implementation

[0013] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided to make the description of this application more complete and comprehensive, and to fully convey the concept of the exemplary embodiments to those skilled in the art. The drawings are merely illustrative illustrations of this application and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted.

[0014] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more exemplary embodiments. Numerous specific details are provided in the following description to give a full understanding of the exemplary embodiments disclosed in this application. However, those skilled in the art will recognize that the technical solutions disclosed in this application can be practiced with one or more specific details omitted, or other methods, components, steps, etc., can be employed. In other instances, well-known structures, methods, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of the disclosure of this application.

[0015] Example 1

[0016] like Figure 1 As shown, this embodiment discloses a keyword search intent analysis method based on GEO, including: S101: Based on the historical data of the target search keywords, analyze semantic fragments and construct an initial clustering structure by using the local deviation features of node density in the contextual co-occurrence patterns between semantic fragments; It should be noted that the semantic fragment specifically includes two parts: basic semantic units and the contextual relationships between basic semantic units.

[0017] Specifically, the basic semantic unit is a semantic expression unit obtained by semantic mapping of historical search data for the target search keywords, and is specifically manifested as a semantic vector based on words or phrases.

[0018] In one specific embodiment, for the user-input target search keyword "apple", basic semantic units are first extracted from historical search data, such as phrases like "Apple phone case", "apple pie recipe", and "apple growing conditions". These basic semantic units are mapped into semantic vectors using pre-trained word embedding models (such as Word2Vec or GloVe), thereby obtaining the semantic representation of each basic semantic unit.

[0019] Furthermore, the contextual association is determined by the conditional co-occurrence probability between basic semantic units and the semantic distance between basic semantic units.

[0020] Specifically, the conditional co-occurrence probability is obtained by counting the number of co-occurrences of basic semantic units in historical search data. For example, for two basic semantic units A and B, if the number of search records in which they co-occur is N(A,B), and the number of search records in which basic semantic unit B appears alone is N(B), then the co-occurrence probability of basic semantic unit A under the condition of basic semantic unit B is calculated as follows: .

[0021] Furthermore, the semantic distance is used to measure the degree of semantic difference between basic semantic units. Specifically, let the semantic vector corresponding to basic semantic unit A be... The semantic vector corresponding to the basic semantic unit B is The semantic distance between two basic semantic units can be calculated using the cosine distance or Euclidean distance between vectors. For example, when using Euclidean distance, it can be expressed as: .

[0022] Furthermore, the conditional co-occurrence probability and semantic distance jointly determine the connection weights of the contextual association relationships between basic semantic units, and the specific calculation method is as follows: ,in, The contextual association connection weights between basic semantic units A and B The semantic distance is calculated above. The exponential function is used to adjust the impact of semantic distance on connection weights; the smaller the semantic distance value, the larger the connection weight.

[0023] A semantic connection structure is formed based on the aforementioned connection weights. Specifically, basic semantic units are used as nodes, and the connection weights are used as edge weights between nodes, thereby forming a semantic connection structure for subsequent analysis.

[0024] For example, regarding the keyword "apple," the basic semantic units "Apple phone case" and "Apple charger" have a higher connection weight because they are semantically similar and have a high probability of conditional co-occurrence; while "Apple phone case" and "apple growing conditions" have a low probability of conditional co-occurrence and a large semantic distance, thus having a lower connection weight. The connection weight calculation process can be accurately reproduced according to the above implementation method.

[0025] Specifically, constructing the initial clustering structure includes: Extract the contextual relationships between basic semantic units to form a semantic connection structure; It should be noted that this process calculates the connection weights of contextual relationships by using the conditional co-occurrence probability and semantic distance of basic semantic units in historical data. Then, it constructs a semantic connection structure using basic semantic units as nodes and connection weights as edge weights. Specifically, the semantic connection structure is a weighted directed graph structure, where each node represents a basic semantic unit, the edges between nodes represent the contextual relationships between basic semantic units, and the weights on the edges are the connection weights. .

[0026] Furthermore, it should be understood that the connection weights between basic semantic units are determined in the following way: First, calculate the conditional co-occurrence probability of basic semantic unit A with other basic semantic unit B from historical search data, denoted as . Secondly, the semantic distance between basic semantic unit A and basic semantic unit B is calculated, denoted as . The semantic distance is calculated using the Euclidean distance or cosine distance of the semantic vectors of the basic semantic units. The specific calculation method has been disclosed in the above content and will not be repeated here. For example, assuming the conditional co-occurrence probability of the basic semantic units "Apple phone case" and "Apple charger" is 0.65, and the semantic distance between them is calculated to be 0.15, then the connection weight can be calculated as follows: ; By calculating the above methods among all basic semantic units, a complete semantic connection structure can be constructed. The weighted graph structure contains all nodes and the connection relationships and weights between corresponding nodes.

[0027] Based on the coupling offset between the node connection distribution of the basic semantic unit in the semantic connection structure and the connection distribution of the corresponding neighboring nodes, the local offset state of each basic semantic unit is determined. It should be understood that the local offset state is used to describe the difference in the connection distribution between the basic semantic unit and its neighboring nodes in the semantic connection structure, so as to further identify the local deviation features of node density.

[0028] The determination of the local offset state of each basic semantic unit includes: Calculate the offset between the node connection distribution of each basic semantic unit and the average connection distribution of its corresponding neighboring nodes; In practical implementation, the connection distribution of the basic semantic unit node A is first determined, that is, the set of connection weights between node A and all its directly connected nodes, denoted as a set. , where nodes First, identify the neighboring nodes directly connected to node A. Second, calculate the average connection distribution of node A's neighboring nodes, which is the average value of the connection weights of each neighboring node to other nodes. Specifically, if neighboring nodes The set of connection weights with its neighboring nodes is Then the neighboring nodes The average connectivity distribution is: ; Furthermore, the overall average connectivity distribution of the neighboring nodes of node A is the average of the average connectivity distributions of each neighboring node: ; Then, the offset between the node connection distribution of node A and the overall average connection distribution of its neighboring nodes is calculated as the absolute value of the difference between the two: ; For example, taking the basic semantic unit "Apple phone case" as an example, assuming its average connection distribution is 0.72 and the overall average connection distribution of its neighboring nodes is 0.54, then the offset of this node is: .

[0029] Based on the coupling relationship between the offset and the semantic distance between the basic semantic units, the local offset state of each basic semantic unit is determined. Specifically, the local offset state can be represented as the degree of local density deviation of a node. Local offset state of a node. It is determined by both the offset and the semantic distance to neighboring nodes, and is calculated as follows: ,in, The offset is the value calculated in the preceding steps. Represents node A and its neighboring nodes The semantic distance between them is used, and the exponential function is used to assign the influence of the semantic distance on the local offset state.

[0030] Based on the local offset state of each basic semantic unit, the spatial positions are non-uniformly rearranged to generate the initial clustering structure; It should be understood that this process aims to adjust the relative positions of each basic semantic unit in the semantic space based on the local offset states of each basic semantic unit obtained above, so as to highlight the local deviation features and thus provide a reliable basis for subsequent clustering analysis.

[0031] Specifically, the non-uniform rearrangement method of spatial positions includes the following processing: First, determine the position coordinates of each basic semantic unit in the initial semantic space. It should be noted that the initial position coordinates of each basic semantic unit are obtained through its corresponding semantic vector representation. Specifically, the semantic vector representation of each basic semantic unit is obtained through a semantic embedding model (such as Word2Vec, GloVe, BERT, etc.) built from historical search data, and serves as the initial position coordinates. Second, based on the local offset state values ​​determined above... The position coordinates of each basic semantic unit are adjusted. Specifically, the magnitude of the position coordinate adjustment is proportional to the local offset state value of the corresponding basic semantic unit, and the adjustment direction is to migrate from the region with a smaller local offset state value to the region with a larger local offset state value. Furthermore, the specific formula for adjusting the position coordinates is as follows:

[0032] in, The adjusted position coordinates of the basic semantic unit A. The coordinates are the original position coordinates. This is the local offset state of the basic semantic unit A. Let be the initial position coordinates of the neighboring node Bi. This represents the local offset state of the neighboring node Bi. To adjust the amplitude coefficient, it is determined through experiments or data verification, with a specific value range, for example, between 0.1 and 1.

[0033] For example, taking the basic semantic unit "Apple phone case" as an example, if the local offset state value of the basic semantic unit... The calculated value is 0.18, and the local offset state values ​​of neighboring nodes such as "Apple charger" and "Apple screen protector" are 0.25 and 0.20 respectively. The initial semantic space position coordinates are vector coordinates: Apple phone case: ; Apple charger: ; Apple phone screen protector: .

[0034] Furthermore, after adjusting the position according to the above formula, the new spatial coordinates of the "Apple phone case" can be obtained.

[0035] Finally, based on the adjusted spatial coordinates of all basic semantic units, a clustering algorithm is used to generate an initial clustering structure; Specifically, spatial density clustering algorithms, such as the DBSCAN clustering algorithm, can be used for clustering, taking the adjusted coordinates of the basic semantic units as input; clustering parameters (such as neighborhood radius) can be used. The minimum number of samples (MinPts) can be determined by analyzing the density distribution of basic semantic units obtained from historical search data, in order to ensure the generation of a reasonable clustering structure.

[0036] For example, after obtaining the initial clustering structure through the above process, related semantic units such as "Apple phone case", "Apple charger", and "Apple phone screen protector" will cluster into an independent clustering region, while agricultural or food-related semantic units such as "Apple planting method" and "Apple pie making method" will form another independent clustering region, so as to clearly reflect the differences in intent of the search keywords.

[0037] S102: Based on the local sparse distribution features of the semantic matching relationship between the initial cluster structure nodes and keywords, extract the local abnormal node region of the semantic matching relationship; It should be noted that the semantic matching relationship refers to the degree of semantic relevance between each basic semantic unit and the target search keyword, which is obtained by calculating the semantic mapping relationship between the basic semantic unit and the target search keyword.

[0038] Specifically, the semantic matching relationship includes the following two aspects: First, the semantic mapping relationship between basic semantic units and target search keywords; Second, the semantic matching strength distribution generated based on the above semantic mapping relationship.

[0039] Furthermore, the specific calculation method for the above semantic mapping relationship includes: First, obtain the semantic vectors corresponding to each basic semantic unit and the target search keyword. The specific vector acquisition method is as described in S101. The same semantic embedding model (such as BERT model or Word2Vec model) can be used to obtain the semantic vector representation. Then, the semantic mapping relationship between the basic semantic unit vector and the target search keyword vector is determined by calculating the semantic projection error between them.

[0040] It should be understood that the semantic projection error is calculated by the cosine distance or Euclidean distance between the basic semantic unit vector and the target keyword vector. The specific calculation method, either cosine similarity or Euclidean distance, can be selected based on the actual data characteristics.

[0041] For example, taking the keyword "apple" and the basic semantic unit "Apple phone case" as examples, their semantic vector representations are obtained respectively: The semantic vector for the keyword "apple" is: ; The semantic vector of the basic semantic unit "Apple phone case" is: ; The semantic projection error between the two (exemplarily calculated using cosine distance) is: ; Specifically, the distribution of semantic matching strength between basic semantic units and keywords can be determined based on the projection error value obtained from the semantic mapping relationship.

[0042] Furthermore, based on the semantic matching intensity distribution corresponding to all basic semantic units, the local sparse distribution features of the semantic matching relationship between the nodes of the initial cluster structure and the keywords are determined; Specifically, the local sparse distribution feature refers to the phenomenon that the semantic matching strength is lower in a specific local region within the initial clustering structure than in other regions; it can be quantitatively described using statistical indicators such as the mean or median of the local matching strength.

[0043] For example, in the initial clustering structure, if the local region node set is {"Apple planting method", "Apple cultivation technique", "Apple fertilization method"}, the semantic matching strength (CosSim) values ​​of its match with the keyword "Apple" are 0.15, 0.18 and 0.12 respectively, which are significantly lower than those of other region nodes (e.g., "Apple phone case" is 0.85 and "Apple charger" is 0.83). This local region exhibits a sparse distribution of semantic matching strength.

[0044] Specifically, the extraction of local abnormal node regions in semantic matching relationships includes: The discrete matching state of each basic semantic unit is calculated based on the semantic matching intensity distribution; It should be understood that this process determines the discrete matching state of the basic semantic unit in its region by quantitatively analyzing the difference in semantic matching strength between each basic semantic unit and the clustering neighborhood nodes.

[0045] Specifically, the discrete state matching calculation process includes: First, for each basic semantic unit in the initial clustering structure, calculating the semantic matching strength between the basic semantic unit and the target keyword; Second, determining the set of neighborhood nodes of the basic semantic unit in the initial clustering structure; wherein, the determination of neighborhood nodes can adopt the same neighborhood radius parameter used in the clustering analysis process (such as the neighborhood radius in the DBSCAN clustering algorithm). Then, based on the difference between the semantic matching strength between the basic semantic unit and the target keyword and the matching strength of the nodes in the neighborhood node set, the matching discrete state of the basic semantic unit is calculated. Specifically, the calculation of matching discrete states uses statistical indicators of the degree of dispersion, such as the coefficient of variation or standard deviation. For example, consider a basic semantic unit A and its set of neighboring nodes. For example, let the semantic matching strength of the basic semantic unit A be . The semantic matching strengths of the neighboring nodes are respectively Then the matching discrete state of the basic semantic unit A It can be calculated as follows: First, calculate the average matching strength of the neighboring node set: Then, calculate the matching discrete state values ​​(e.g., expressed as absolute differences): ; It should be understood that this discrete state value A higher value indicates a greater difference in semantic matching relationship between the basic semantic unit A and the set of neighboring nodes, and vice versa.

[0046] Based on the coupling offset relationship between the matching discrete state and the local connection distribution of the initial cluster structure where the basic semantic unit is located, identify the basic semantic units whose matching discrete state and connection distribution are inconsistent. It should be understood that the purpose of this process is to identify nodes where there is an inconsistency between the semantic matching strength and the local connectivity features of the initial clustering structure; these nodes are potential local outliers.

[0047] The basic semantic unit for identifying and matching discrete states that are inconsistent with the connection distribution includes: Calculate the difference between the matching discrete state of each basic semantic unit and the matching discrete state of its neighboring nodes; Specifically, for each basic semantic unit, the difference between the matching discrete states of the basic semantic unit and its neighboring nodes is calculated. For example, the basic semantic unit A and the set of neighboring nodes The matching discrete states are respectively The formula for calculating the difference is: Furthermore, the overall discrete dissimilarity of the basic semantic unit A is obtained. For example, take the average of all differences: .

[0048] Based on the coupling relationship between the difference and the semantic distance between the basic semantic units, the basic semantic units that match the discrete states and the connection distribution are determined. It should be noted that this processing is based on the joint analysis of the difference between the discrete states of the matching and the semantic distance between the basic semantic units, in order to determine whether there are semantic matching anomalies in a specific basic semantic unit in a local region.

[0049] Specifically, for the set of neighboring nodes of a basic semantic unit, the semantic distance between each neighboring node and the target basic semantic unit is calculated.

[0050] It should be understood that the semantic distance is calculated in the same way as the semantic vector distance above, and is usually calculated using cosine distance or Euclidean distance.

[0051] Furthermore, based on the joint relationship between the difference obtained from the above calculation and the corresponding semantic distance, basic semantic units with inconsistencies between the matching discrete states and the connection distribution are identified; Specifically, threshold-based judgment methods or local anomaly detection methods can be used to identify basic semantic units that have inconsistent matching discrete states and connection distributions.

[0052] For example, taking basic semantic unit A as an example, the set of neighborhood nodes is: The calculated differences are respectively The corresponding semantic distances are respectively Then the following judgment rules can be adopted: If the difference between a certain neighboring node Greater than the set difference threshold At the same time, corresponding semantic distance Less than the set distance threshold ,Right now: If so, it is determined that there is an inconsistency between the matching discrete state and the connection distribution between the basic semantic unit A and its neighboring nodes; Furthermore, if the basic semantic unit A and multiple neighboring nodes all satisfy the above conditions, then the basic semantic unit A is determined to be an inconsistent basic semantic unit.

[0053] The identified basic semantic units are aggregated according to spatial adjacency to generate local abnormal node regions with semantic matching relationships; Specifically, after identifying the inconsistent basic semantic units, these basic semantic units are further aggregated into regions according to their spatial location and adjacency relationship in the initial clustering structure to obtain local abnormal node regions of semantic matching relationship; It should be understood that the spatial adjacency relationship here refers to the direct connection relationship between nodes in the initial clustering structure. If two inconsistent basic semantic units have a direct connection path in the clustering structure, then they will be aggregated into the same region. For example, if both basic semantic unit A and basic semantic unit B are determined to be inconsistent basic semantic units, and nodes A and B are directly connected or indirectly connected through an intermediate node in the initial clustering structure, then A and B are aggregated into the same abnormal node region.

[0054] Furthermore, through the above-mentioned judgment of spatial adjacency relationships and node aggregation operation, one or more local abnormal node regions with semantic matching relationships are finally obtained.

[0055] S103: Based on the asymmetric topological features of the context path between nodes within the local abnormal node region, reconstruct the semantic path between nodes to form an optimized set of candidate semantic paths. Specifically, the reconstructed semantic path between nodes includes: Extract the context path connections between basic semantic units within the region of local anomaly nodes; It should be understood that the purpose of this process is to clarify the specific semantic path connection relationships between basic semantic units within the local abnormal node region in order to obtain the original path structure for subsequent reconstruction.

[0056] Specifically, for each local abnormal node region, all basic semantic units within that region are first obtained; Furthermore, based on the co-occurrence sequence of semantic units in the original historical search data, the contextual path connection relationship between basic semantic units within the local abnormal node region is extracted to form an initial path set; For example, suppose a certain local anomaly node region contains basic semantic units. Path sequences that frequently appear in historical search data are These sequences are then extracted as contextual path connections within the abnormal region. It should be noted that the extracted context path connection relationship is represented in the form of a directed graph. Each basic semantic unit corresponds to a node in the directed graph, and the semantic units in each path correspond to the edge connection direction in the directed graph. By constructing the local directed graph structure, scattered abnormal nodes can be reorganized into a semantic flow with temporal characteristics, providing a topological basis for subsequent mining of asymmetric intent pointers.

[0057] Calculate the coupling offset of the forward path connection relationship and the backward path connection relationship between basic semantic units to determine the asymmetric connection structure in the path direction; Specifically, the purpose of this process is to identify the differences in connection directions between different basic semantic units in the semantic path in order to determine whether there are asymmetric features in the path structure, thereby determining the basis for subsequent path reconstruction. In one specific embodiment, for any two basic semantic units within a local abnormal node region, such as basic semantic unit A and basic semantic unit B, the forward path frequency from A to B in the original historical search data is statistically analyzed. and the backward path frequency from B to A. ; Furthermore, the coupling offset of the path connection relationship is calculated based on the path frequency values ​​obtained from statistics in order to determine the asymmetric connection structure in the path direction; For example, the coupling offset of the path connection relationship It can be defined as: When the coupling offset When the value is small, it indicates that the path connection is relatively symmetrical in the forward and backward directions; conversely, when the coupling offset is small... When the value is large, it indicates that the path connection relationship has asymmetric characteristics; Furthermore, after performing the above calculation on the coupling offset of all basic semantic unit pairs, basic semantic unit pairs with asymmetric connection relationships are selected to determine the asymmetric connection structure in the path direction. The context path connections between basic semantic units are rearranged based on the asymmetric connection structure to generate an optimized set of candidate semantic paths; It should be understood that this processing aims to optimize and adjust the original extracted contextual path connection relationships based on the asymmetric connection structure obtained above in the path direction, so as to obtain a set of candidate semantic paths that better reflects the actual search behavior.

[0058] Specifically, for basic semantic unit pairs that have been determined to have an asymmetric connection structure, the priority in the connection direction is determined based on the statistical results of the above coupling offsets.

[0059] For example, for a basic semantic unit pair (A, B), if the statistically obtained forward path frequency... Greater than the backward path frequency Therefore, when reordering paths, priority should be given to maintaining the connection order from basic semantic unit A to basic semantic unit B, and excessive introduction of reverse connection paths should be avoided. Furthermore, based on the aforementioned priority adjustment strategy, the original context path connections within the local abnormal node region are analyzed and adjusted one by one: For example, the original set of paths also contains path sequences. and If the path sequence is determined based on the above statistical analysis If the connection has a higher priority, it will be retained during the rearrangement process. For reverse sequences Then, it should be weakened or removed appropriately.

[0060] Through the above-described sequential rearrangement operation, an optimized set of candidate semantic paths that can more accurately reflect the user's actual search behavior pattern is obtained. After the above rearrangement process is completed, the optimized candidate semantic path set can be reconstructed into a directed graph structure representation to facilitate subsequent processing.

[0061] S104: Combining the non-stationary characteristics of the state transition topology of the context sequence in the candidate semantic path set, the subdivided search intent corresponding to the target keyword is determined by inferring the path state through the conditional random field model; Specifically, the purpose of this process is to further analyze the optimized candidate semantic path set and use the Conditional Random Field (CRF) model to infer the state of the context sequence composed of each basic semantic unit, thereby accurately identifying the specific search intent of the target search keyword; Specifically, the path state inference through the conditional random field model includes: The path connection order of basic semantic units is determined based on the candidate semantic path set, and a path state transition sequence is constructed. Specifically, based on the optimized candidate semantic path set mentioned above, the specific connection order between basic semantic units within each path is extracted to form a path state transition sequence for input to the conditional random field model. It should be noted that each candidate path corresponds to a path state transition sequence, which uses basic semantic units as basic state nodes, and the connection order between each state node corresponds to the transition order in the sequence. For example, if the candidate semantic path set contains a path sequence of "Apple phone → phone case → protective case", then the state transition sequence corresponding to this path is represented as: ; Furthermore, after performing the above processing on all candidate semantic paths one by one, the complete path state transition sequence is obtained to construct the path state set; The aforementioned path state set is used as the input sequence for the conditional random field model, providing the model with basic structured feature data; Calculate the connection offset between adjacent states in the state transition sequence to determine the state transition boundary; Specifically, to make the path state inference more accurate, the connection offset between adjacent states in the state transition sequence is further calculated to identify the actual state transition boundaries within the sequence. It should be noted that the connection offset in this process refers to the difference in transition probabilities between two adjacent states, which is used to determine the state switching position during the transition process. In one specific embodiment, for any adjacent state pair (X, Y) within the state transition sequence, the actual transition probability from state X to state Y in historical data is statistically analyzed. and the reverse transition probability And use this to calculate the connection offset. Specifically: ; It should be understood that when the connection offset When the value is small, it indicates that the directionality between state transitions is weak and it is not used as a state transition boundary; conversely, if the connection offset is large... If the value is large, then it is determined that there is a state transition boundary between adjacent states (X, Y). Furthermore, through the above calculation and analysis of connection offsets, multiple state transition boundaries in the path state transition sequence are identified; For example, if the state transition sequence is "iPhone → Phone Case → Protective Case", and the calculated connection offset of "iPhone → Phone Case" is small while the connection offset of "Phone Case → Protective Case" is large, then the state transition boundary is determined to be located at the position of "Phone Case → Protective Case", indicating that there is a significant change in semantic state during the actual state transition of the sequence from "Phone Case" to "Protective Case". Based on the state transition boundary, the path state of the state transition sequence is inferred using a conditional random field model to generate the detailed search intent corresponding to the target search keyword. It should be understood that the purpose of this process is to use the conditional random field model to infer the path state transition sequence with a defined state transition boundary, thereby accurately identifying the specific search intent of the target keyword. Specifically, the path state transition sequence is used as the input sequence of the conditional random field model, each basic semantic unit is used as the state node of the CRF model, and the transition probability between adjacent state nodes is used as the transition feature function of the CRF model to perform probabilistic inference of the path state. It should be noted that the CRF model can be trained based on historical search data and is used to predict the search intent label corresponding to each basic semantic unit of a given state transition sequence. Furthermore, for each path state transition sequence, after prediction using the CRF model, each basic semantic unit will obtain a corresponding state label, and each label corresponds to a specific sub-category of search intent. For example, given the state transition sequence {iPhone → Phone Case → Protective Case}, the labels corresponding to each state node after inference by the CRF model can be: Apple iPhone: Purchase intention for electronic products; Phone case: Intended purchase of electronic product accessories; Protective case: Intended purchase of electronic product accessories; Furthermore, based on the distribution of state labels and state transition boundaries obtained from the CRF model, the search intent labels of the entire path sequence are integrated to determine the specific sub-search intent corresponding to the target keyword. For example, in the above example path, since the state transition boundary is located at "phone case → protective case", it means that the search intent remains consistent from this position. Therefore, the overall search intent of the path can be determined as "purchase intent for electronic product accessories", thus accurately reflecting the user's actual search purpose. After performing the above CRF model state inference and segmented search intent integration processing on all optimized candidate semantic paths, a complete set of segmented search intents corresponding to the target search keywords is finally obtained to achieve accurate search intent analysis.

[0062] Example 2

[0063] like Figure 2 As shown in the example, the parts not detailed in this embodiment are as shown in Example 1. This embodiment discloses a keyword search intent analysis system based on GEO, including: Module 201 is used to parse semantic fragments based on historical data of target search keywords and construct an initial clustering structure through the local deviation features of node density in the context co-occurrence pattern between semantic fragments. Extraction module 202 is used to extract the local sparse distribution features of the semantic matching relationship between the initial cluster structure nodes and keywords, and to extract the local abnormal node regions of the semantic matching relationship. The optimization module 203 is used to reconstruct the semantic path between nodes based on the topological asymmetric features of the context path between nodes within the local abnormal node region, thereby forming an optimized candidate semantic path set. The determination module 204 is used to combine the non-stationary features of the state transition topology of the context sequence in the candidate semantic path set and determine the subdivided search intent corresponding to the target keyword through path state inference of the conditional random field model.

[0064] 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, weights, and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0065] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A keyword search intent analysis method based on GEO, characterized in that, include: Based on historical data of the target search keywords, semantic fragments are analyzed, and an initial clustering structure is constructed using the local deviation features of node density in the contextual co-occurrence patterns among semantic fragments; including: Extract the contextual relationships between basic semantic units to form a semantic connection structure; Based on the coupling offset between the node connection distribution of the basic semantic unit in the semantic connection structure and the connection distribution of the corresponding neighboring nodes, the local offset state of each basic semantic unit is determined. Based on the local offset state of each basic semantic unit, the spatial positions are non-uniformly rearranged to generate the initial clustering structure; Based on the local sparse distribution features of semantic matching relationships between nodes in the initial cluster structure and keywords, local abnormal node regions of semantic matching relationships are extracted. Based on the asymmetric topological features of the contextual paths between nodes within the local abnormal node region, the semantic paths between nodes are reconstructed to form an optimized set of candidate semantic paths; including: Extract the context path connections between basic semantic units within the region of local anomaly nodes; Calculate the coupling offset of the forward path connection relationship and the backward path connection relationship between basic semantic units to determine the asymmetric connection structure in the path direction; The context path connections between basic semantic units are rearranged based on the asymmetric connection structure to generate an optimized set of candidate semantic paths; By combining the non-stationary characteristics of the state transition topology of the context sequence in the candidate semantic path set, and through path state inference using a conditional random field model, the detailed search intent corresponding to the target keyword is determined; including: The path connection order of basic semantic units is determined based on the candidate semantic path set, and a path state transition sequence is constructed. Calculate the connection offset between adjacent states in the state transition sequence to determine the state transition boundary; Based on the state transition boundary, the path state of the state transition sequence is inferred using a conditional random field model, generating the detailed search intent corresponding to the target search keywords.

2. The method according to claim 1, characterized in that, The semantic fragment includes basic semantic units and the contextual relationships between basic semantic units. The basic semantic units are obtained by semantic mapping from the historical search data of the target search keywords. The contextual relationships are determined by the conditional co-occurrence probability between basic semantic units and the semantic distance between basic semantic units.

3. The method according to claim 1, characterized in that, Determining the local offset state of each basic semantic unit includes: Calculate the offset between the node connection distribution of each basic semantic unit and the average connection distribution of its corresponding neighboring nodes; Based on the coupling relationship between the offset and the semantic distance between the basic semantic units, the local offset state of each basic semantic unit is determined.

4. The method according to claim 1, characterized in that, The semantic matching relationship includes the semantic mapping relationship between basic semantic units and target search keywords, as well as the semantic matching intensity distribution generated based on the semantic mapping relationship, wherein the semantic mapping relationship is determined by the semantic projection error between the basic semantic units and the target search keywords.

5. The method according to claim 4, characterized in that, The extraction of local abnormal node regions for semantic matching relationships includes: The discrete matching state of each basic semantic unit is calculated based on the semantic matching intensity distribution; Based on the coupling offset relationship between the matching discrete state and the local connection distribution of the initial cluster structure where the basic semantic unit is located, identify the basic semantic units whose matching discrete state and connection distribution are inconsistent. The identified basic semantic units are aggregated according to spatial adjacency to generate local abnormal node regions with semantic matching relationships.

6. The method according to claim 5, characterized in that, The basic semantic unit for identifying and matching discrete states that do not match the connection distribution includes: Calculate the difference between the matching discrete state of each basic semantic unit and the matching discrete state of its neighboring nodes; Based on the coupling relationship between the difference and the semantic distance between the basic semantic units, the basic semantic units that match the discrete states and the connection distribution are determined.

7. A keyword search intent analysis system based on GEO, implemented using the method described in any one of claims 1-6, characterized in that, include: The building module is used to parse semantic fragments based on historical data of target search keywords and construct an initial clustering structure through the local deviation features of node density in the contextual co-occurrence patterns between semantic fragments. The extraction module is used to extract the local sparse distribution features of the semantic matching relationship between the nodes of the initial cluster structure and the keywords, and to extract the local abnormal node regions of the semantic matching relationship. The optimization module is used to reconstruct the semantic paths between nodes based on the asymmetric topological features of the context paths between nodes within the local abnormal node region, thereby forming an optimized set of candidate semantic paths. The determination module is used to combine the non-stationary features of the state transition topology of the context sequence in the candidate semantic path set, and determine the subdivided search intent corresponding to the target keyword through path state inference by the conditional random field model.

Images