Method, device and equipment for tapping tactics of a kind of sports based on natural language interaction and storage medium

Abstract

The application discloses a shooting sport tactic mining method based on natural language interaction, and comprises the following steps: obtaining a multivariate event sequence set; splitting an original tactic set or a current tactic set in the multivariate event sequence based on a global constraint term to obtain a plurality of single values, randomly combining the plurality of single values into a plurality of tactics, and iteratively combining the plurality of tactics to obtain a global candidate tactic set; a fine-tuning generator is used to adjust the single values in the tactics selected by experts based on a local constraint term to obtain a fine-tuning candidate tactic set; natural language is converted into the global constraint term or the local constraint term based on an NLP module, a global constraint objective function is constructed based on the global constraint term, a mining optimizer is used to obtain an optimal global candidate tactic, a fine-tuning optimizer is used to replace the selected tactics with the fine-tuning candidate tactic to construct a new tactic set, and an optimal fine-tuning candidate tactic is obtained when the value of the objective function is the smallest. The method can convert the natural language of experts and further mine the natural language to obtain a more satisfactory tactic method.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of sports data analysis, and particularly relates to a racket sport tactic mining method and device based on natural language interaction, equipment and storage medium. BACKGROUND

[0002] Racket sports refer to sports such as table tennis and badminton. The characteristic of racket sports is that players need to use rackets to hit the ball alternately on both sides of the net. In such sports, the basic data unit is a hit. A series of continuous alternating hits from one side serving to one side scoring is called a rally. When hitting the ball each time, the player often considers multiple hit details, including the positions of the players, the current position of the ball, the hitting technique, the rotation of the ball, and the like. Different racket sports consider similar parts of the hit details (such as the positions of the players, the position of the ball, the hitting technique, etc.), but not completely the same (such as badminton often does not consider rotation, but considers the height of the ball). In each rally, the player can use various tactics by using the changes in the hit details. Tactics is a high-level countermeasure. For example, the classic left-right mobilization tactic is composed of three hits: the player first hits the ball to the far right of the opponent (the position of the ball); the opponent runs to the right to return the ball (the position of the player); the player again uses an offensive hitting technique (the hitting technique) to quickly hit the ball to the far left of the opponent (the position of the ball), making it difficult for the opponent to return the ball. Tactics is an important point for analysis in the field of racket sports.

[0003] Event sequence data is a common data type: an event is a data record of an event subject, occurrence time, and behavior type (i.e., who did what at what time); an event sequence is a series of related events (such as all events of a subject in a day) arranged in chronological order. An important concept in event sequence is feature pattern, i.e., frequently occurring subsequence. The data of racket sports is essentially an event sequence: each hit can be regarded as an event; each rally can be regarded as an event sequence; and the tactics frequently used by the player can be regarded as a feature pattern. However, racket sports is a special event sequence: the behavior type of each event contains the records of multiple hit details. We call this event sequence a multi-element event sequence. As shown in FIG. 1, the tactic t is used in two sequences s1 and s2, and can be split into several single values (sv1-sv4). Figure 1

[0004] ​Sports tactic mining technology refers to finding the tactics (i.e., feature pattern set) of several players from a large number of round data (i.e., a multi-element event sequence data set). However, traditional sports tactic mining technology is data-driven, that is, it can find the feature pattern with the highest frequency of occurrence, but it does not care about the actual analysis value of the mining result. For example, in table tennis, the subsequence with the highest frequency of occurrence is the mutual attack of the two sides (i.e., the players of the two sides alternately use offensive hitting techniques), but such a subsequence does not have a tactical analysis value, because it is known to all coaches, and coaches are more concerned about the selection of the ball route (i.e., the change of the landing point of the ball on the table) when attacking. In this case, the analysis expert hopes to be able to give suggestions for modification to the algorithm mining result, and hopes that the algorithm can further subdivide the attack tactic according to the change of the landing point to divide it into several different tactical systems.

[0005] J. Wang, J. Wu, A. Cao, Z. Zhou, H. Zhang, and Y. Wu. Tac-miner: Visual tactic mining for multiple table tennis matches. IEEE Transactions on Visualization and Computer Graphics, 27(6):2770-2782, 2021. doi:10.1109 / TVCG.2021.3074576 discloses a technology for visual mining analysis of tactics in multiple table tennis matches, which does not support detailed adjustment of the mining result. SUMMARY

[0006] The present application provides a racket sports tactic mining method based on natural language interaction, which can convert the expert's natural language and further mine a more satisfactory tactic method.

[0007] A racket sports tactic mining method based on natural language interaction, comprising:

[0008] (1) obtaining a multi-element event sequence set;

[0009] (2) constructing a mining generator and a fine-tuning generator, the mining generator is used to split the original tactic set or the current tactic set in the multi-element event sequence based on global constraints to obtain a plurality of single values, randomly combine the plurality of single values into a plurality of tactics, and iteratively combine the plurality of tactics to obtain a global candidate tactic set; the fine-tuning generator is used to adjust the single value in the tactic selected by the expert based on the local constraint term to obtain a fine-tuning candidate tactic set;

[0010] (3) Constructing a mining optimizer and a fine-tuning optimizer, the expert gives a suggestion, i.e. natural language, for the original tactic set or the current tactic set each time, the natural language is converted into a global constraint term or a local constraint term based on the NLP module, a global constraint objective function is constructed based on the global constraint term, the mining optimizer is used to minimize the value of the global constraint objective function to obtain the best global candidate tactic, the fine-tuning optimizer is used to replace the selected tactic with a fine-tuning candidate tactic to construct a new tactic set, and the new tactic set is substituted into the global constraint objective function, and the best fine-tuning candidate tactic is obtained when the value of the objective function is minimized.

[0011] The natural language of the expert is one-to-one corresponding to the constraint term through the NLP module, and the constraint term includes a global constraint term and a local constraint term.

[0012] The global constraint term includes a tactic position constraint term, a tactic length constraint term and a detail importance constraint term, wherein the tactic position constraint term is:

[0013]

[0014] The tactic length constraint term is:

[0015]

[0016] The detail importance constraint term is:

[0017] imp(v i )=c

[0018] Wherein, t i ≤s m Indicates that the tactic t i is used in the multi-element event sequence s m , i is the index of the tactic, m is the index of the multi-element event sequence, c is a constant of -1 to 1, and is the importance set by the user, -1 indicates unimportant, and 1 indicates important.

[0019] The local constraint term includes an ignored tactic constraint term, a detailed tactic detail constraint term, an extracted tactic core constraint term, an expanded tactic context constraint term, a simplified tactic constraint term and a replaced tactic detail constraint term.

[0020] The global constraint objective function L * (S,T) is:

[0021]

[0022]

[0023] Wherein, S is a multi-element event sequence set, T is a global candidate tactic set, freq(t i ) is the i-th tactic t iFrequency of use, sv(v j ,s m ) represents the m-th multivariate event sequence s m The j-th shot detail v j The corresponding single value cannot be expressed as the number of tactics; L(S,T) is a baseline optimization objective, α and β are polynomial-order balance coefficients, and sv(s m ) represents the m-th multivariate event sequence s m The single values ​​in the equation cannot be expressed as the number of tactics. `idx_con(.)`, `len_con(.)`, and `imp(.)` are the tactical position constraint, tactical length constraint, and detail importance constraint, respectively. The purpose of the polynomial order-of-magnitude balancing coefficient is to balance the order of magnitude of these three constraints. For example, suppose there are 1000 sequences, each with 10 single values, and 20 tactics are found in the sequences. Then, the first term in L(S,T) is 20, the second term is on the order of several hundred (a tactic might be used in several hundred sequences), and the third term is on the order of several thousand (each sequence will have several single values, accumulating to several thousand). These three terms are completely different in magnitude, making it difficult to ensure the balance of optimization when optimizing L(S,T) (i.e., the third term will be compressed to the maximum extent, while the first two terms are basically not considered). The polynomial order-of-magnitude balancing coefficient can be used to achieve this optimization balance.

[0024] The fine-tuning optimizer is used to replace the selected tactic with a fine-tuning candidate tactic to construct a new tactical set T. new for:

[0025] T new =T-adjT+{c∈canT}

[0026] Where adjT is the selected tactic, c is the candidate tactic for fine-tuning, and canT is the set of candidate tactics for fine-tuning.

[0027] A device for mining tactics of clapping sports based on natural language interaction, comprising:

[0028] Database, used to obtain multivariate event sequence sets;

[0029] The front-end module provides a suggestion, i.e., natural language, for each original or current tactical set. The NLP module then transforms the natural language into global or local constraints.

[0030] The back-end module is used for splitting the original or current tactic set in the multi-element event sequence into a plurality of single values based on global constraints, randomly combining the plurality of single values into a plurality of tactics, and iteratively combining the plurality of tactics to obtain a global candidate tactic set; used for adjusting the single values in the selected tactics based on local constraint terms to obtain a fine-tuned candidate tactic set; constructing a global constraint objective function based on the global constraint terms, using a global optimizer to minimize the global constraint objective function value to obtain an optimal global candidate tactic, and using a fine-tuning optimizer to replace the selected tactics with the fine-tuned candidate tactic to construct a new tactic set, and substituting the new tactic set into the global constraint objective function, and obtaining an optimal fine-tuned candidate tactic when the objective function value is minimized.

[0031] A computer device comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the natural language interaction-based racket sport tactic mining method when executing the computer program.

[0032] A computer readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the natural language interaction-based racket sport tactic mining method when executed by a processor.

[0033] Compared with the prior art, the beneficial effects of the present application are:

[0034] The present application converts the natural language of experts into constraint terms by using an NLP module, and constructs a generator and an optimizer based on the constraint terms, and continuously iterates to mine tactics based on the generator and the optimizer, so that satisfactory tactics are obtained. It can be used for the tactic analysis of other professional racket players. Players can analyze their own tactical characteristics and further improve their competitive level; players can also analyze the characteristics of the opponent's playing method. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The data structure diagram provided for the background art;

[0036] Figure 2 The natural language interaction-based racket sport tactic mining device flowchart provided for the specific implementation;

[0037] Figure 3 The natural language interaction-based racket sport tactic mining method flowchart provided for the specific implementation;

[0038] Figure 4 The visual user interface diagram provided for the specific implementation, wherein, Figure 4 (a) is a visual user interface layout diagram, Figure 4 (b) is a visual diagram of each part of the visual user interface;

[0039] Figure 5 Partial constraint term schematic diagram provided for the specific embodiment. Specific embodiment

[0040] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0041] The present application provides a kind of based on natural language interaction's shot sports tactics mining method, as shown in Figure 3 Specific steps are as follows:

[0042] (1) obtain a set of multivariate event sequences, i.e.

[0043] (2) build mining generator and fine-tuning generator, mining generator is used to split the original tactic set or the current tactic set in the multivariate event sequence based on global constraint term to obtain multiple single values, randomly combine the multiple single values into multiple tactics, and iteratively combine the multiple tactics to obtain a global candidate tactic set; fine-tuning generator is used to add, delete and replace the single value in the tactic selected by the expert to obtain a fine-tuning candidate tactic set;

[0044] (3) build mining optimizer and fine-tuning optimizer, the expert gives a suggestion each time for the original tactic set or the current tactic set, i.e. natural language, based on NLP module, the natural language is converted into global constraint term or local constraint term, the present application adopts NLP module based on template. Based on a large number of natural language modification suggestions collected in expert interview, a rich natural language template (147 templates) is established this time. For example, the expert input "the tactic needs to analyze the subsequent development" will match the template "<tactic id> needs to analyze <analysis direction> development", and the corresponding parameter t i And direction. The corresponding relationship between the natural language of the expert and the constraint term is shown in Table 1.

[0045] Table 1 corresponding relationship between natural language of expert and constraint term

[0046]

[0047] Constraint term is as follows:

[0048] 1. Set tactical position: Different sports, different players, different tactical sequence. For example, table tennis is more concerned about the service tactics, badminton is more concerned about the last few strokes. Therefore, experts need to set the position of the tactics. This constraint is achieved by adding a loss term idx_con(t ) to the optimization objective: This function counts the percentage of tactical t

[0049] i each time the sequence of strokes exceeds the expert's expectations. When a certain tactic is mostly used outside the expert's expected range, the function value is larger, causing the algorithm to prefer to delete the tactic to reduce the optimization objective. The tactical position constraint term is:

[0050] Where t i represents the tactical t m . m i is used in the sequence of events s m .

[0051] 2. Set tactical length: Because longer tactics often have more information, but the frequency of occurrence will be less; while shorter tactics often have less information, but the frequency is high. In order to balance between the two, experts will need to adjust the length of the tactics. This constraint is achieved by adding a loss term len_con(t i ) to the optimization objective: The value of this function is 0 by default, but when the length of the tactic is not within the expert's expected range, the value is 1, prompting the algorithm to delete the tactic.

[0052] The tactical length constraint term is:

[0053]

[0054] 3. Set the importance of details: Because different players have different tactical emphasis, the analysis also needs to pay attention to specific details of the strokes. This constraint is achieved by setting the importance of each stroke detail imp(v i ) and adjusting the coefficient of the importance to change the proportion of the stroke detail in the tactic. The detail importance constraint term is:

[0055] imp(v i ) = c

[0056] Where c is a constant between -1 and 1, and is the importance set by the user, -1 means not important, 1 means important.

[0057] 4. Ignore tactics: For some useless tactics, experts want to delete directly. This constraint will directly delete the tactic t i .

[0058] 5. Make tactical details explicit: For some ambiguous tactics, experts would like to make some important values explicit, or even split the tactic into several tactics according to different values. The constraint t i →sub(t i ,values) will replace tactic t i with several possible sub-tactics, where values mark the positions of values that need to be made explicit (these values are originally "empty"). As shown in Fig. Figure 5 , when the user wants to make the hitting property v2 of event e3 in tactic t explicit (which is originally empty), the algorithm needs to find all possible single values that the empty value could correspond to in the original sequence s that uses tactic t, and try to insert them into the empty value to generate several candidate tactics. For example, the algorithm finds that sequence s uses tactic t at position z, so it fills in the empty value to generate the fine-tuned result.

[0059] 6. Extract the core of a tactic: For many complicated tactics, experts would like to extract the core hitting details, or even merge several similar tactics. The constraint t i →sup(t i ,values) will replace tactic t i with a super-tactic, where values mark the hitting details that need to be made "empty". This constraint is the inverse of constraint 5, as shown in Fig. Figure 5 , when the user wants to ignore the hitting property v1 of event e2 in tactic t (which is originally not empty), the algorithm needs to replace the specific single value with an empty value to generate a candidate tactic.

[0060] 7. Extend the context of a tactic: Experts would like to extend the pre- or post- shots of some tactics to complete the full process of the tactic. The constraint t i →exp(t i ,direction) will replace tactic t i with an extended tactic, where direction represents the extension direction (forward or backward). As shown in Fig. Figure 5 , when the user wants to extend the subsequent development of tactic t, the algorithm needs to find all possible single values that could appear in the hitting events after using tactic t in the original sequence s, and add them to the end of tactic t to generate several candidate tactics. For example, the algorithm finds that there are two single values in e5 after sequence s uses tactic t, so the algorithm will generate two candidate tactics, adding c and z to the end of tactic t respectively (only the case of adding c is shown in the figure, and the case of adding z is similar), to generate the fine-tuned results.

[0061] 8. Simplify a tactic: This constraint is just the opposite of the last constraint, as shown in Fig. Figure 5 , when the user wants to simplify the first shot of a tactic, the algorithm needs to directly delete the first shot to generate a candidate tactic.

[0062] 9. Replace Tactical Details: Experts may modify a specific shot detail within a tactic. The constraint `tirep(ti, original value, target value)` will change a specific shot detail (original value) in tactic `ti` to a specific value (target value). For example... Figure 5 As shown, when a user wants to replace the shot attribute v1 (originally c) of time e2 in tactic t with a, the algorithm needs to perform a replacement operation to generate a candidate tactic.

[0063] A global constraint objective function is constructed based on global constraint terms. The mining optimizer is used to minimize the value of the global constraint objective function to obtain the best global candidate tactic. The fine-tuning optimizer is used to replace the selected tactic with the fine-tuning candidate tactic to construct a new tactic set. The new tactic set is substituted into the global constraint objective function. When the objective function value is minimized, the best fine-tuning candidate tactic is obtained.

[0064] Global constraint objective function L * (S,T) is:

[0065]

[0066]

[0067] Where S is a multivariate event sequence set, T is a global candidate tactic set, and freq(t) i That is, the i-th t t i Frequency of use, sv(v j ,s m ) represents the m-th multivariate event sequence s m The jth shot detail v j The corresponding single value cannot be expressed as the number of tactics; L(S,T) is a baseline optimization objective, α and β are polynomial-order balance coefficients, and sv(s m ) represents the m-th multivariate event sequence s m The single value in the formula cannot be expressed as the number of tactics. idx_con(.), len_con(.), and imp(.) are the tactical position constraint, tactical length constraint, and detail importance constraint, respectively.

[0068] The fine-tuning optimizer is used to replace the selected tactic with a fine-tuning candidate tactic to construct a new tactical set T. new for:

[0069] T new =T-adjT+{c∈canT}

[0070] Wherein, adjT is the selected tactic, c is the fine-tuning candidate tactic, and canT is the fine-tuning candidate tactic set.

[0071] A tactic mining device based on natural language interaction, as shown in Figure 2 , comprising:

[0072] A database for obtaining a multi-element event sequence set;

[0073] A front-end module, wherein an expert gives a suggestion for the original tactic set or the current tactic set each time, i.e., natural language, and the natural language is converted into a global constraint term or a local constraint term based on an NLP module;

[0074] A back-end module, wherein the original tactic set or the current tactic set in the multi-element event sequence is split based on the global constraint to obtain a plurality of single values, the plurality of single values are randomly combined into a plurality of tactics, and the plurality of tactics are iteratively combined to obtain a global candidate tactic set; the single value in the selected tactic is adjusted based on the local constraint term to obtain a fine-tuning candidate tactic set; a global constraint objective function is constructed based on the global constraint term, a mining optimizer is used to minimize the global constraint objective function value to obtain an optimal global candidate tactic, a fine-tuning optimizer is used to replace the selected tactic with the fine-tuning candidate tactic to construct a new tactic set, and the new tactic set is substituted into the global constraint objective function, and the optimal fine-tuning candidate tactic is obtained when the objective function value is minimized.

[0075] Visual design, as shown in Figure 4 (a) and (b), and the specific steps are:

[0076] As shown in A of Figure 4 , the control bar allows the user to select the data set, player, and opponent of interest, and filters out the corresponding original sequence set for analysis.

[0077] As shown in B of Figure 4 , the query view, the suggestion panel allows the user to input natural language suggestions, and displays the converted constraints for the user to check and correct. The suggestion panel also records the history operation, and the user can backtrack to the previous history state or undo the last modification.

[0078] As shown in E of Figure 4 , the projection view projects each tactic into a two-dimensional polar coordinate system according to similarity. The polar angle encodes the most relevant hitting details of the tactic; the polar radius encodes the similarity of the tactic: the technology uses the PCA algorithm to project the tactic onto a one-dimensional length as the polar radius. The size of each projected point encodes the frequency of the tactic, and the color encodes the win rate of the tactic. When previewing and adjusting, as shown in C1 and E1 of Figure 4 , the tactic list and the projection view will use plus and minus signs to mark the added and deleted tactics, respectively.

[0079] As shown in Figure 4 The tactics list lists all the tactics and displays the core information of each tactic, such as usage rate, win rate, etc. Each tactic is displayed by several patterns, each of which encodes the hitting details of one frame. Each icon is composed of several parts, each of which encodes a hitting detail. For null values in the tactics, the corresponding part in the pattern will use a semi-transparent icon to display the most frequent value and its frequency ratio. When previewing adjustments, the tactics list will top the deleted / added tactics and use minus / plus signs to mark them.

[0080] As shown in Figure 4 The round view allows the user to browse the specific use of a tactic in each round after selecting it. The round view includes three parts: the uppermost bar chart displays the position statistics of each use of the tactic; the middle round list lists the specific hitting detail values of each round; the user can click any row in the round list to trigger the video below to play the corresponding game segment.

Claims

1. A method for mining tactical patterns in clapping movements based on natural language interaction, characterized in that, include: (1) Obtain a multivariate event sequence set; (2) Construct a mining generator and a fine-tuning generator. The mining generator is used to split the original tactical set or the current tactical set in the multi-element event sequence based on global constraints to obtain multiple single values, randomly combine multiple single values ​​into multiple tactics, and iteratively combine multiple tactics to obtain a global candidate tactical set. The fine-tuning generator is used to adjust the single values ​​in the tactics selected by experts based on local constraints to obtain a fine-tuning candidate tactical set. (3) Construct mining optimizer and fine-tuning optimizer. Experts give suggestions, i.e. natural language, for the original tactic set or the current tactic set each time. The natural language is converted into global or local constraints based on the NLP module. A global constraint objective function is constructed based on the global constraint. The mining optimizer is used to minimize the value of the global constraint objective function to obtain the best global candidate tactic. The fine-tuning optimizer is used to replace the selected tactic with the fine-tuning candidate tactic to construct a new tactic set. The new tactic set is substituted into the global constraint objective function. When the objective function value is minimized, the best fine-tuning candidate tactic is obtained. Global constraints include tactical position constraints, tactical length constraints, and detail importance constraints. The tactical position constraints are as follows: The tactical length constraint is: The detail importance constraint is as follows: ,in, Indicating tactics t i In multi-event sequence s m It is used in China. i As an index for tactics, m An index for a multivariate event sequence. c A constant ranging from -1 to 1, representing the user-defined importance level, where -1 indicates no importance and 1 indicates importance; global constraint objective function. for: Where S is a multivariate event sequence set, and T is a global candidate tactic set. freq ( t i That is, the first i One tactic t i Frequency of use sv ( v j ,s m ) represents the first m A multi-dimensional event sequence s m The Middle j individual shot details v j The corresponding single value cannot be expressed as the number of tactics; L(S,T) is a baseline optimization objective. , These are the polynomial order of magnitude balance coefficients. For the first m A multi-dimensional event sequence s m The single value in the text cannot be expressed as the number of tactics.

2. The method for mining tactical patterns of clapping movements based on natural language interaction according to claim 1, characterized in that, The NLP module maps the expert's natural language to constraints, including global and local constraints.

3. The method for mining tactical patterns of clapping movements based on natural language interaction according to claim 1, characterized in that, Local constraints include ignoring tactical constraints, specifying tactical details constraints, extracting core tactical constraints, expanding tactical context constraints, simplifying tactical constraints, and replacing tactical detail constraints.

4. The method for mining tactical patterns of clapping movements based on natural language interaction according to claim 1, characterized in that, The fine-tuning optimizer is used to replace the selected tactic with a fine-tuning candidate tactic to construct a new tactical set T. new for: ,in, To select tactics, To fine-tune the candidate tactics, To fine-tune the candidate tactic set.

5. A device for mining tactics of clapping sports based on natural language interaction, employing the method for mining tactics of clapping sports based on natural language interaction as described in any one of claims 1-4, characterized in that, include: Database, used to obtain multivariate event sequence sets; The front-end module provides a suggestion, i.e., natural language, for each original or current tactical set. The NLP module then transforms the natural language into global or local constraints. The backend module is used to split the original or current tactical set in a multi-element event sequence into multiple individual values ​​based on global constraints, randomly combine multiple individual values ​​into multiple tactics, and iteratively combine multiple tactics to obtain a global candidate tactical set; it is used to adjust the individual values ​​in the tactics selected by experts based on local constraints to obtain a fine-tuned candidate tactical set; it constructs a global constraint objective function based on global constraints, mines an optimizer to minimize the value of the global constraint objective function to obtain the best global candidate tactic, and fine-tunes an optimizer to replace the selected tactic with the fine-tuned candidate tactic to construct a new tactical set, substitutes the new tactical set into the global constraint objective function, and obtains the best fine-tuned candidate tactic when the objective function value is minimized; Global constraints include tactical position constraints, tactical length constraints, and detail importance constraints. The tactical position constraints are as follows: The tactical length constraint is: The detail importance constraint is as follows: ,in, Indicating tactics t i In multi-event sequence s m It is used in China. i As an index for tactics, m An index for a multivariate event sequence. c A constant ranging from -1 to 1, representing the user-defined importance level, where -1 indicates no importance and 1 indicates importance; global constraint objective function. for: Where S is a multivariate event sequence set, and T is a global candidate tactic set. freq ( t i That is, the first i One tactic t i Frequency of use sv ( v j ,s m ) represents the first m A multi-dimensional event sequence s m The Middle j individual shot details v j The corresponding single value cannot be expressed as the number of tactics; L(S,T) is a baseline optimization objective. , These are the polynomial order of magnitude balance coefficients. For the first m A multi-dimensional event sequence s m The single value in the text cannot be expressed as the number of tactics.

6. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for mining tactical movements based on natural language interaction as described in any one of claims 1-4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is processed and executed, it implements the steps of the method for mining tactical movements based on natural language interaction according to any one of claims 1-4.

Images