Sports event stream real-time parsing method
By constructing an event association graph G based on temporal offset feature Δt and semantic similarity feature S, and combining topology reconstruction and path search, the problems of temporal drift and semantic conflict of multi-source heterogeneous data are solved, and high-precision real-time parsing of sports event streams is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI XUANTI INFORMATION TECHNOLOGY CO LTD
- Filing Date
- 2026-04-28
- Publication Date
- 2026-07-24
Smart Images
Figure CN122454482A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of event data processing technology, specifically to a method for real-time parsing of sports event streams. Background Technology
[0002] In high-concurrency live sports broadcasts, event stream data (such as goals, fouls, and substitutions) is typically generated in real-time from multi-source heterogeneous signals (video frame annotations, sensor data, and manual annotations). However, within millisecond-level time windows, there are micro-level time drifts and semantic granularity inconsistencies between different sources. This is especially true in complex events occurring consecutively within a very short period (such as a goal being disallowed after an offside), where event overlap, out-of-order parsing, or duplicate parsing can easily occur. This leads to imperceptible but fine-grained errors in downstream statistical and referee assistance systems, affecting the accuracy of judgments. Therefore, how to perform refined and sequential parsing of event streams under ultra-low latency conditions has become a pressing technical problem that needs to be solved. Summary of the Invention
[0003] The purpose of this invention is to provide a real-time parsing method for sports event streams to address the shortcomings of the prior art.
[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for real-time parsing of sports event streams, comprising: Obtain the raw event stream data E to be parsed, and perform timestamp normalization on E to obtain the event sequence E0 under a unified timeline; E0 is divided into sliding windows, the candidate event set in each window is extracted, and the temporal offset feature Δt and semantic similarity feature S between events are calculated to obtain the feature set F; Construct an event association graph G based on F, where nodes represent events and edge weights are obtained by weighting Δt and S to form initial event relationships; Perform topology reconstruction under global constraints on G, identify and resolve implicit conflicts through minimum conflict path search and local consistency backtracking, and form the basis for event sequence rearrangement; Based on the event sequence rearrangement criteria, the event state convergence determination and structured output are performed to generate a final event stream with continuous semantic closure, so as to achieve real-time parsing.
[0005] Preferably, calculating the temporal offset features between events includes: Based on the event sequence E0 under a unified time axis, the standardized timestamps of any adjacent event pairs in the candidate event set are extracted to construct an initial time difference sequence. The time difference is then corrected by combining the initial time difference sequence with the historical delay distribution of the data source to obtain a corrected time difference sequence. Local smoothing is performed on the corrected time difference sequence to suppress the interference of sudden jitter on the time relationship and form a stable time change trend. A directional time series offset feature is constructed based on the smoothed time change trend.
[0006] Preferably, calculating semantic similarity features includes: performing structured semantic encoding on each event in the candidate event set, mapping event type, participating objects, and behavioral attributes to a unified semantic representation vector, and constructing an event semantic expression sequence; based on the event semantic expression sequence, extracting contextual association information between events, enhancing the semantic representation of adjacent events to obtain a semantically enhanced representation containing contextual dependencies; performing multi-dimensional matching analysis on the semantically enhanced representation of any event pair, calculating the semantic closeness from two levels: semantic consistency and behavioral evolution correlation, forming an initial semantic similarity relationship set; and performing global consistency calibration on the initial semantic similarity relationship set to generate stable semantic similarity features.
[0007] Preferably, after calculating the temporal offset feature Δt and the semantic similarity feature S, the two types of features are combined according to event pairs to construct a feature set F, where each element contains the temporal offset feature Δt and the semantic similarity feature S of the corresponding event pair, and maintains the index relationship consistent with the event order in the original event sequence E0.
[0008] Preferably, the construction of the event association graph model G based on F includes: Using each event in the event sequence E0 under a unified time axis as a node, an initial node set is established according to the index relationship of the events in the feature set F, and the temporal offset feature Δt and semantic similarity feature S corresponding to each node are retained; Secondly, for any pair of nodes, a composite edge weight is constructed based on the temporal offset feature Δt and the semantic similarity feature S, and then combined according to a preset ratio coefficient to obtain the weighted connection relationship between the nodes. Based on the weighted connection relationship, node pairs that meet the connection threshold conditions are selected, directed edges are established to form an initial association structure, and redundant connections are eliminated by limiting the maximum number of outgoing edges of a single node, thus obtaining the event association graph G.
[0009] Preferably, the steps of performing topology reconstruction under global constraints on the event association graph G and identifying and resolving latent conflicts include: Traverse all directed paths in the event association graph G, construct a path cost sequence based on edge weight accumulation, and filter the candidate path set that meets the conditions by combining time monotonicity constraints and semantic consistency constraints; perform minimum conflict path search in the candidate path set, and select the path with the minimum overall conflict cost as the main path structure by comparing the fluctuation amplitude of edge weight and the number of direction changes in the path. Based on the main path structure, local consistency backtracking is performed on its neighboring nodes. Nodes in the unselected paths that have temporal reversal or semantic deviation from the main path are detected layer by layer and reconnected or removed to resolve implicit conflicts. The topology of the event association graph G is updated according to the connection relationship after conflict resolution, forming the basis for reordering the event sequence that satisfies global consistency constraints.
[0010] Preferably, the overall structure of the event association graph G is updated, including: deleting all connection edges that are determined to be in conflict, replacing the edge relationships after reconnection, and recalibrating the number of outgoing edges of each node so that it still satisfies the constraint that each node retains 3 outgoing edges.
[0011] Preferably, generating the final event stream with continuous semantic closure includes: Based on the event sequence rearrangement criteria, the nodes in the event association graph are rearranged in order to generate candidate event sequences. The continuity of the temporal offset feature Δt and semantic similarity feature S of adjacent events is detected. When the absolute value of the temporal offset feature Δt of adjacent events is less than 0.08 seconds and the semantic similarity feature S is greater than 0.65, they are marked as continuous evolution states. For event segments marked as continuously evolving states, state convergence determination is performed. By statistically analyzing the fluctuation range of semantic similarity feature S within the event segment, the state convergence interval is identified when the difference between the maximum and minimum values is less than 0.2. Semantic fusion processing is performed on events within the state convergence interval to merge multiple events into a single structured event unit, while retaining its start and end times and dominant semantic tags. The fused structured event unit is then integrated with the non-converged events in chronological order to generate the final event stream, thereby enabling real-time parsing and output of sports event events.
[0012] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention introduces an event association graph G constructed by synergistically driving temporal offset feature Δt and semantic similarity feature S. Based on this, it combines topological reconstruction under global constraints, minimum conflict path search, and local consistency backtracking to optimize event relationships. Existing technologies typically rely solely on temporal sorting or single semantic matching to linearly concatenate events, making it difficult to handle fine-grained temporal drift and implicit semantic conflicts in multi-source heterogeneous data. This invention, however, constructs an event association graph G with directional causal relationships from the bottom layer by directionally mapping the temporal offset feature Δt and segmenting and modulating the semantic similarity feature S, and then dynamically adjusting the weights of both based on event density. This transforms the association between events from simple adjacency relationships into a structured relationship network that can express potential triggering mechanisms. This technique directly solves the problems of event misordering and weak association interference in high-concurrency scenarios, resulting in higher discrimination accuracy and stability of initial event relationships.
[0013] 2. This invention employs a topology reconstruction strategy combining minimum conflict path search and local consistency backtracking on the event association graph G to proactively identify and resolve implicit conflicts. Furthermore, it introduces an event state convergence determination and semantic fusion output mechanism. Unlike existing post-processing methods based on fixed rules or simple threshold filtering, this invention quantifies edge weight fluctuations and direction changes through path conflict costs and performs layer-by-layer backtracking correction on neighboring nodes using the main path structure as the core, ensuring global consistency constraints in the conflict resolution process. Simultaneously, by determining the semantic stability of continuously evolving event segments and performing structured fusion, semantic closure is achieved directly at the data structure level, thus avoiding fragmented output. These technical means ensure that the final generated event stream is strictly consistent in time sequence and continuous and complete in semantic expression, significantly improving the accuracy and usability of real-time parsing results, especially in complex and composite event scenarios, effectively reducing misjudgments and information redundancy. Attached Figure Description
[0014] 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.
[0015] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.
[0017] For examples, please refer to Figure 1 As shown in this embodiment, the real-time parsing method for sports event streams includes: Obtain the original event stream data E to be parsed, and perform timestamp standardization on E to obtain the event sequence E0 under a unified timeline.
[0018] First, the raw event stream data E to be parsed is acquired. This raw event stream data E originates from various heterogeneous data channels, including but not limited to event annotation data output by the video recognition system, sensor-collected data, and manually entered event data. Due to differences in the acquisition frequency, time base, and transmission delay of different data sources, the timestamps carried by each event exhibit fine-grained offsets and inconsistencies.
[0019] To address this, the present invention standardizes the timestamps in the original event stream data E. Specifically, a unified reference timeline is constructed, mapping the local time of each data source onto this reference timeline. During the mapping process, a time correction mechanism based on data source characteristics is introduced to adaptively compensate for time offsets from different sources, eliminating the effects of systematic errors and random jitter. Simultaneously, by combining the event arrival order with a buffer queue, out-of-order events are reordered to ensure they meet timing consistency constraints under a unified timeline.
[0020] During the timestamp standardization process, a time consistency identifier is added to each event to characterize the stability and reliability of its time location, thus providing a foundation for subsequent event correlation analysis. Through this process, an event sequence E0 is obtained under a unified timeline, ensuring comparability and continuity of events in the time dimension, providing reliable input for subsequent feature extraction and correlation modeling.
[0021] The event sequence E0 under a unified time axis is divided into sliding windows, and candidate event sets are extracted in each time window. At the same time, the temporal offset feature Δt and semantic similarity feature S between events are calculated to form a feature set F.
[0022] Specifically, the length and step size of the sliding window are first determined based on the distribution density of the event sequence E0 on the time axis. The window length is dynamically set according to the statistical results of the number of events per unit time. When the number of events per unit time is higher than 20, the window length is set to 0.5 seconds to avoid information aliasing caused by high-density event stacking; when the number of events is lower than 20, the window length is extended to 1.5 seconds to ensure the integrity of the event context. The step size is set at 0.5 times the window length to maintain a 50% overlap between adjacent windows, thereby ensuring that events are not interrupted by window boundaries.
[0023] Based on the aforementioned window length and step size, a continuously covered time window sequence is generated on a unified time axis.
[0024] Within each time window, corresponding local event subsequences are extracted according to the chronological order of event timestamps, and the candidate event set is filtered based on event temporal proximity and semantic label consistency. Specifically, the filtering rules are as follows: when the time interval between two events is less than 0.3 times the window length and the event type labels belong to the same semantic category, they are grouped into the same candidate event set; for events with a time interval greater than the above threshold but exhibiting a continuous triggering relationship, their event trigger identifier fields are checked for dependencies, and these events are included in the same candidate event set.
[0025] Using the above method, several candidate event sets are formed within each time window, and the original sequence information of events within each set is preserved.
[0026] After obtaining the candidate event set, the time-series offset feature Δt is calculated for any adjacent event pair within the set. The specific process is as follows: First, the standardized timestamps of adjacent events are extracted, and the time difference between them is calculated as the initial time difference. Then, the initial time difference is corrected by combining historical latency statistics from different data sources. Historical latency is obtained by statistically analyzing the average latency and variance of the past 1000 events. When the latency fluctuation of a data source exceeds twice its average, the corresponding time difference is offset to compensate for its return to the stable latency range of that data source. After completing the time difference correction, the continuous time difference sequence is smoothed using a sliding smoothing process. The smoothing window length is set to three event pairs, and the median of the time difference within the window is used to replace the current value to suppress abnormal fluctuations caused by sudden jitter. Finally, based on the smoothed time difference sequence and the event occurrence sequence identifier, a directional time-series offset feature Δt is generated. Δt is positive when the timestamp of the later event is greater than that of the earlier event, and negative otherwise, to characterize the sequential relationship between events and the potential triggering direction.
[0027] Simultaneously, semantic similarity features S are calculated for events in the candidate event set.
[0028] First, the event type, participating objects, and behavioral attributes of each event are mapped to a fixed-length 128-dimensional semantic representation vector. The event type is represented using one-hot encoding, the participating objects are mapped to numerical vectors through a pre-defined entity encoding table, and the behavioral attributes are converted into continuous vectors using a word vector model. These three types of vectors are then concatenated to form a unified semantic representation vector. Next, event neighborhood relationships are constructed within the candidate event set. For each event, the two closest events before and after it in time are selected as neighborhood events, and their semantic representation vectors are weighted and fused. Neighboring events that are closer to the current event have higher weights. The weights are calculated using the inverse distance normalization method, so that the sum of all neighborhood weights is 1, thus obtaining a semantically enhanced representation.
[0029] Based on this, similarity is calculated for the semantically enhanced representation of any event pair. Specifically, the semantic closeness is represented by calculating the cosine value of the angle between the two vectors. When the cosine value is close to 1, it indicates that the semantics are highly consistent, and when it is close to 0, it indicates that the semantics are unrelated. Finally, the similarity results of all event pairs are normalized and mapped to the interval between 0 and 1 to form a stable semantic similarity feature S.
[0030] After calculating the temporal offset feature Δt and the semantic similarity feature S, the two types of features are combined according to event pairs to construct a feature set F, where each element contains the temporal offset feature Δt and the semantic similarity feature S of the corresponding event pair, and maintains the index relationship consistent with the event order in the original event sequence E0.
[0031] The feature set F can simultaneously reflect the temporal dependency and semantic association between events, providing a unified and quantifiable input basis for subsequent event association modeling and conflict resolution.
[0032] Construct an event association graph G based on F, where nodes represent events and edge weights are obtained by weighting Δt and S to form initial event relationships.
[0033] First, a node set is constructed based on the event sequence E0 under a unified time axis. Each event in the event sequence E0 is mapped to a unique node in the event association graph G. Each node is assigned an identifier number according to the order index of the events in the feature set F, ensuring a one-to-one correspondence between nodes and event pairs in the feature set F. Internally, each node is also associated with all the temporal offset features Δt and semantic similarity features S that it participated in constructing, so that they can be directly called upon in subsequent edge weight calculations.
[0034] After the node set is constructed, candidate connections are established for the relationship between any two nodes. For any node pair, its corresponding temporal offset feature Δt is extracted first, and the temporal sequence is determined according to the direction of Δt. When Δt is positive, it indicates that the previous event points to the next event in time, and only directed connections from the previous node to the next node are allowed. When Δt is negative, connections in the opposite direction are established. When the absolute value of Δt is less than 0.02 seconds, it is considered an approximately synchronous event in time. Such node pairs are allowed to be connected bidirectionally in subsequent processing, but their edge weights need to be adjusted separately. The aforementioned value of 0.02 seconds is based on the minimum resolvable time interval of high-speed actions (such as shooting or touching the ball) in sports events. This interval is less than 1 / 50th of a second of the regular video frame interval, which can effectively distinguish between real continuous actions and acquisition errors.
[0035] After determining the connection direction, the temporal offset feature Δt is processed using interval mapping. Specifically, the absolute value of Δt is divided into three intervals: when Δt is less than 0.1 seconds, it is considered a strong temporal correlation interval, and is assigned a high weight value between 0.8 and 1; when Δt is between 0.1 seconds and 0.5 seconds, it is considered a medium correlation interval, and the corresponding weight decreases linearly from 0.8 to 0.3; when Δt is greater than 0.5 seconds, it is considered a weak correlation interval, and the corresponding weight is limited to between 0 and 0.3, gradually approaching 0 as the time interval increases. This method ensures that events that are closer in time have a higher correlation strength, while guaranteeing that events with longer time intervals do not interfere with the graph structure.
[0036] Simultaneously, the semantic similarity feature S is modulated in segments. Specifically, when the semantic similarity feature S is greater than 0.75, an exponential enhancement strategy is adopted, that is, non-linear amplification is applied to the original value to further widen the differences within the 0.75 to 1 range, thereby highlighting the semantic connection between highly related events; when the semantic similarity feature S is between 0.4 and 0.75, its original value is kept unchanged to reflect a moderate semantic association; when the semantic similarity feature S is less than 0.4, it is compressed to map it to the 0 to 0.2 range to reduce the impact of weak semantic associations on the overall graph structure. The above threshold settings are based on the statistical results of the event label distribution in actual event data, where the semantic similarity of highly related events is usually concentrated above 0.75, while that of irrelevant events is generally below 0.4.
[0037] After obtaining the temporal correlation strength corresponding to the temporal offset feature Δt and the semantic modulation weight corresponding to the semantic similarity feature S, the two are combined according to a preset ratio coefficient.
[0038] The preset proportional coefficient is not a fixed value, but is dynamically adjusted based on the event density within the time window of the current node. When the number of events within the time window is greater than 30, it indicates that events occur frequently. In this case, the weight ratio of the temporal offset feature Δt is set to 0.7, and the weight ratio of the semantic similarity feature S is set to 0.3 to strengthen the dominant role of temporal order in relation determination. When the number of events is less than or equal to 30, the weight ratios of the two are adjusted to 0.5 and 0.5 respectively to balance temporal and semantic information in sparse scenarios. The composite edge weight of each node pair is calculated in the above manner, and the value of the composite edge weight is limited to between 0 and 1.
[0039] After calculating the composite edge weight, all node pairs are filtered for connections. Specifically, the edge weight threshold is set to 0.35. When the composite edge weight of a node pair is greater than or equal to 0.35, its corresponding directed connection is retained; when the composite edge weight is less than 0.35, the node pair is considered to lack a stable association and no connection is established.
[0040] This threshold is obtained through statistical analysis of the event association accuracy in historical event data. Under this threshold, false connections can be effectively suppressed while ensuring recall.
[0041] After filtering the edges, the number of outgoing edges for each node is further limited.
[0042] Specifically, each node is allowed a maximum of 3 outgoing edges, which are sorted from highest to lowest weight. Only the 3 connections with the highest weights are retained, and the rest are deleted. This restriction is designed to avoid generating too many redundant paths in high-density event scenarios, thereby improving the structural sparsity and resolvability of the event association graph G.
[0043] Through the above steps, an event association graph G is finally formed, where nodes represent events and nodes are connected by directional edges. The edge weights are obtained by weighting the temporal offset feature Δt and the semantic similarity feature S.
[0044] The event association graph G can simultaneously express the temporal dependencies and semantic relationships between events in a unified structure, thereby forming initial event relationships and providing a foundation for subsequent conflict detection and path optimization.
[0045] Perform topology reconstruction under global constraints on G, identify and resolve implicit conflicts through minimum conflict path search and local consistency backtracking, and form the basis for event sequence rearrangement.
[0046] First, all directed paths in the event association graph G are traversed. During the traversal, each node is used as the starting point, and the graph expands layer by layer along its outgoing edges, generating a set of paths from the starting node to the ending node. During path expansion, the composite edge weight of each edge in the path is cumulatively calculated to form a path cost sequence. This path cost sequence is formed by summing the composite edge weights of all edges in the path in sequence, while introducing a path length normalization rule, which divides the cumulative value by the number of edges in the path, to avoid bias in cost evaluation due to excessively long paths.
[0047] After the path cost sequence is formed, temporal monotonicity constraints and semantic consistency constraints are introduced to filter the paths. The temporal monotonicity constraint requires that the temporal offset feature Δt of all adjacent nodes in the path maintains the same direction, that is, it is continuously positive or has no more than one sign change within the allowable range, and the absolute value of Δt corresponding to the sign change is less than 0.05 seconds. This threshold is derived from the temporal jitter range of brief synchronous actions in actual competitions.
[0048] The semantic consistency constraint calculates the average semantic similarity feature S of all edges in the path. A path is considered semantically consistent if the average is greater than 0.6; otherwise, it is considered semantically deviating and eliminated. This dual constraint process yields a set of candidate paths that meet the criteria.
[0049] Perform a minimum conflict path search in the candidate path set.
[0050] Specifically, the degree of conflict is quantified by calculating the fluctuation amplitude of edge weights and the number of direction changes in the path. The fluctuation amplitude of edge weights is obtained by summing the absolute values of the differences between adjacent edge weights in the path. The larger this value is, the more unstable the internal correlation of the path is. The number of direction changes is calculated by counting the number of times the sign of the temporal offset feature Δt changes in the path.
[0051] The two indicators mentioned above are combined in a ratio of 0.6 and 0.4 to form the path conflict cost value. All candidate paths are then ranked, and the path with the lowest conflict cost value is selected as the main path structure. This ratio is set based on the actual characteristics of the continuity of the event, where edge weight stability has a greater impact on path reliability than direction change.
[0052] After determining the main path structure, local consistency backtracking is performed on its neighboring nodes, with the main path structure as the core.
[0053] Starting from each node in the main path, a reverse traversal is performed to its unselected neighboring nodes to detect the relationship between these nodes and the main path. For each neighboring node, firstly, it is determined whether its temporal offset feature Δt with the corresponding node in the main path satisfies the time monotonicity requirement. When there is a Δt sign opposite to the direction of the main path and the absolute value is greater than 0.1 seconds, it is determined that the node has a time reversal conflict. Secondly, the semantic similarity feature S between the node and the main path nodes is calculated. When this value is less than 0.5, it is determined that there is a semantic deviation problem.
[0054] For detected latent conflict nodes, a differentiated processing strategy is implemented. When a node has a temporal reversal conflict but a semantic similarity feature S greater than 0.7, it is reconnected, that is, its edge direction in the event association graph G is adjusted to be consistent with the main path direction, and its composite edge weight is recalculated. When a node has both temporal reversal conflict and semantic deviation problems, the connection between the node and the main path is directly removed to avoid error propagation. When a node has only a slight semantic deviation and its semantic similarity feature S is between 0.5 and 0.7, it is retained but its edge weight is reduced to 0.5 of the original value to reflect its instability.
[0055] After completing the local consistency backtracking, the overall structure of the event association graph G is updated. Specifically, this involves deleting all edges identified as conflicting, replacing them with reconnected edges, and recalibrating the number of outgoing edges for each node to ensure that each node retains a maximum of 3 outgoing edges. Simultaneously, a path connectivity check is performed again on the updated graph structure to ensure that there are no isolated nodes or broken paths.
[0056] Through the above topology reconstruction process, the event association graph G is transformed from the initial structure into an optimized structure that satisfies the global consistency constraint. In this process, implicit conflicts are effectively identified and resolved, and the connection relationships between nodes are more stable and have clear temporal and semantic orientation.
[0057] The resulting structure provides a reliable basis for subsequent event sequence rearrangement, ensuring that the output event sequence remains consistent in both temporal order and semantic logic.
[0058] Based on the event sequence rearrangement criteria, the event state convergence determination and structured output are performed to generate a final event stream with continuous semantic closure, so as to achieve real-time parsing.
[0059] First, the nodes in the event association graph are rearranged according to the event sequence rearrangement criteria. These criteria originate from the topologically reconstructed event association graph structure. Candidate event sequences are generated by linearly expanding the nodes along the main path structure. During this expansion, connection paths with higher edge weights are preferentially selected as the main sequence, while branch paths with edge weights less than 0.8 times the average of the main path are temporarily stored as auxiliary sequences to avoid weakly associated paths interfering with the continuity of the main sequence. The candidate event sequences generated in this way exhibit preliminary consistency in both temporal order and semantic association.
[0060] After obtaining the candidate event sequence, continuity detection is performed on the temporal offset feature Δt and semantic similarity feature S between adjacent events. The specific detection rule is as follows: for any pair of adjacent events, if the absolute value of the temporal offset feature Δt is less than 0.08 seconds and the semantic similarity feature S is greater than 0.65, the event pair is marked as a continuous evolution state.
[0061] The value of 0.08 seconds is based on the minimum time resolution requirement for continuous actions in sports events, which can cover typical intervals such as the time between touching the ball and passing it; a semantic similarity feature S greater than 0.65 indicates that the event has strong semantic continuity. Through pairwise detection, the candidate event sequence is divided into several continuous evolutionary state segments and discontinuous state segments.
[0062] After completing the continuity labeling, a state convergence determination is performed on all event segments marked as continuously evolving states. Specifically, for each continuously evolving state segment, the semantic similarity feature S value corresponding to all events within that segment is extracted, and the difference between the maximum and minimum values is calculated. When the difference is less than 0.2, the event segment is considered to exhibit a stable convergence trend semantically, meaning that no significant semantic shift has occurred during the event evolution process; when the difference is greater than or equal to 0.2, the event segment is considered to have semantic fluctuations and does not meet the convergence condition. The threshold of 0.2 is derived from the statistical analysis of the semantic similarity distribution in historical event data. Fluctuations within this range usually correspond to different stages of the same event, while exceeding this range indicates a semantic shift.
[0063] For event segments identified as state convergence intervals, semantic fusion processing is performed. The specific fusion method is as follows: First, the event with the highest mean semantic similarity feature S within the event segment is identified as the dominant event, and its event type and behavioral attributes are used as the dominant semantic label after fusion. Then, the minimum and maximum values of the start and end times of all events within the event segment are taken as the time range of the fused structured event unit. Based on this, the semantic information of other events within the event segment is auxiliaryly encoded and recorded as extended attributes to retain detailed information without affecting the dominant semantic expression. Through this fusion, multiple highly consistent continuous events are compressed into a single structured event unit, thereby reducing redundancy and improving the compactness of the expression.
[0064] For event segments that do not meet the state convergence condition, the original event granularity remains unchanged, and only their position in the candidate event sequence is adjusted to conform to the overall temporal order. At the same time, for events in these event segments with a semantic similarity feature S lower than 0.5, their priority in the final output is further reduced, so that they are treated as low-confidence events in subsequent applications.
[0065] After completing the state convergence determination and semantic fusion, all fused structured event units and non-converged events are re-integrated along a unified timeline. During the integration process, timestamps are used as the primary sorting criterion. When there are structured event units with overlapping times, event units with higher mean semantic similarity features (S-values) are prioritized for retention, and low-priority events are adjusted by time offset, with the adjustment range not exceeding 0.03 seconds to avoid time conflicts. This adjustment range is less than the aforementioned continuity detection threshold of 0.08 seconds, thus avoiding disruption of the overall temporal structure.
[0066] Finally, a final event stream with continuous semantic closure is generated. Each event unit in the final event stream has a clear time range, a dominant semantic label, and necessary extended attribute information, while strictly increasing in chronological order and exhibiting no breaks or conflicts in semantic expression. Through the above processing, real-time parsing output of the sports event stream is achieved, enabling the event stream to maintain high accuracy while possessing good structured characteristics, directly serving subsequent statistical analysis and decision support applications.
[0067] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes 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.
Claims
1. A method for real-time parsing of sports event streams, characterized by: include: Obtain the raw event stream data E to be parsed, and perform timestamp normalization on E to obtain the event sequence E0 under a unified timeline; E0 is divided into sliding windows, the candidate event set in each window is extracted, and the temporal offset feature Δt and semantic similarity feature S between events are calculated to obtain the feature set F; Construct an event association graph G based on F, where nodes represent events and edge weights are obtained by weighting Δt and S to form initial event relationships; Perform topology reconstruction under global constraints on G, identify and resolve implicit conflicts through minimum conflict path search and local consistency backtracking, and form the basis for event sequence rearrangement; Based on the event sequence rearrangement criteria, the event state convergence determination and structured output are performed to generate a final event stream with continuous semantic closure, so as to achieve real-time parsing.
2. The real-time parsing method for sports event streams according to claim 1, characterized in that: Calculate the temporal offset features between events, including: Based on the event sequence E0 under a unified time axis, the standardized timestamps of any adjacent event pairs in the candidate event set are extracted to construct an initial time difference sequence. The time difference is then corrected by combining the initial time difference sequence with the historical delay distribution of the data source to obtain a corrected time difference sequence. Local smoothing is performed on the corrected time difference sequence to suppress the interference of sudden jitter on the time relationship and form a stable time change trend. A directional time series offset feature is constructed based on the smoothed time change trend.
3. The real-time parsing method for sports event streams according to claim 1, characterized in that: Calculating semantic similarity features includes: performing structured semantic encoding on each event in the candidate event set, mapping event type, participating objects, and behavioral attributes to a unified semantic representation vector, and constructing an event semantic expression sequence; based on the event semantic expression sequence, extracting contextual association information between events, enhancing the semantic representation of adjacent events to obtain a semantically enhanced representation containing contextual dependencies; performing multi-dimensional matching analysis on the semantically enhanced representation of any event pair, calculating the semantic closeness from two levels: semantic composition consistency and behavioral evolution correlation, forming an initial semantic similarity relationship set; and performing global consistency calibration on the initial semantic similarity relationship set to generate stable semantic similarity features.
4. The real-time parsing method for sports event streams according to claim 3, characterized in that: After calculating the temporal offset feature Δt and the semantic similarity feature S, the two types of features are combined according to event pairs to construct a feature set F, where each element contains the temporal offset feature Δt and the semantic similarity feature S of the corresponding event pair, and maintains the index relationship consistent with the event order in the original event sequence E0.
5. The real-time parsing method for sports event streams according to claim 1, characterized in that: The event association graph model G constructed based on F includes: Using each event in the event sequence E0 under a unified time axis as a node, an initial node set is established according to the index relationship of the events in the feature set F, and the temporal offset feature Δt and semantic similarity feature S corresponding to each node are retained; Secondly, for any pair of nodes, a composite edge weight is constructed based on the temporal offset feature Δt and the semantic similarity feature S, and then combined according to a preset ratio coefficient to obtain the weighted connection relationship between the nodes. Based on the weighted connection relationship, node pairs that meet the connection threshold conditions are selected, directed edges are established to form an initial association structure, and redundant connections are eliminated by limiting the maximum number of outgoing edges of a single node, thus obtaining the event association graph G.
6. The real-time parsing method for sports event streams according to claim 5, characterized in that: The steps for performing topology reconstruction under global constraints on the event correlation graph G and identifying and resolving latent conflicts include: Traverse all directed paths in the event association graph G, construct a path cost sequence based on edge weight accumulation, and filter the candidate path set that meets the conditions by combining time monotonicity constraints and semantic consistency constraints; perform minimum conflict path search in the candidate path set, and select the path with the minimum overall conflict cost as the main path structure by comparing the fluctuation amplitude of edge weight and the number of direction changes in the path. Based on the main path structure, local consistency backtracking is performed on its neighboring nodes. Nodes in the unselected paths that have temporal reversal or semantic deviation from the main path are detected layer by layer and reconnected or removed to resolve implicit conflicts. The topology of the event association graph G is updated according to the connection relationship after conflict resolution, forming the basis for reordering the event sequence that satisfies global consistency constraints.
7. The real-time parsing method for sports event streams according to claim 6, characterized in that: The overall structure of the event association graph G is updated, including: deleting all connected edges that are determined to be in conflict, replacing the edge relationships after reconnection, and recalibrating the number of outgoing edges of each node so that it still satisfies the constraint that each node retains 3 outgoing edges.
8. The real-time parsing method for sports event streams according to claim 1, characterized in that: The generation of the final event stream with continuous semantic closure includes: Based on the event sequence rearrangement criteria, the nodes in the event association graph are rearranged in order to generate candidate event sequences. The continuity of the temporal offset feature Δt and semantic similarity feature S of adjacent events is detected. When the absolute value of the temporal offset feature Δt of adjacent events is less than 0.08 seconds and the semantic similarity feature S is greater than 0.65, they are marked as continuous evolution states. For event segments marked as continuously evolving states, state convergence determination is performed. By statistically analyzing the fluctuation range of semantic similarity feature S within the event segment, the state convergence interval is identified when the difference between the maximum and minimum values is less than 0.
2. Semantic fusion processing is performed on events within the state convergence interval to merge multiple events into a single structured event unit, while retaining its start and end times and dominant semantic tags. The fused structured event unit is then integrated with the non-converged events in chronological order to generate the final event stream, thereby enabling real-time parsing and output of sports event events.