Training methods, devices, equipment, and storage media for game intent recognition models

Abstract

This application provides a training method, apparatus, device, and storage medium for a game intent recognition model. The method includes: acquiring a basic sample set; calling a pre-trained large language model to perform deep semantic parsing on the data in each basic sample in the basic sample set to obtain the full intent annotation information corresponding to each basic sample; concatenating multiple pre-generated instruction templates corresponding to different intent tasks with each basic sample to obtain multiple input samples corresponding to each intent task; decomposing the full intent annotation information corresponding to each basic sample to obtain the intent annotation information corresponding to each input sample; and performing supervised fine-tuning training of the pre-trained lightweight language model for different intent tasks based on each input sample and the intent annotation information corresponding to each input sample to obtain the trained game intent recognition model. This solution can improve the accuracy of the model's game intent recognition and enhance the model's generalization ability.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a training method, apparatus, device, and storage medium for a game intent recognition model. Background Technology

[0002] As the gaming industry becomes increasingly intelligent, natural language dialogue has gradually replaced traditional button and click operations, becoming the core method for deep interaction between players and game systems. It is widely used in core scenarios such as intelligent NPC dialogue, automated task guidance, in-game intelligent customer service, and voice command control. Game dialogue, in particular, is highly scenario-specific, with players expressing themselves in a highly colloquial and non-standard manner, employing a large amount of game-specific slang, abbreviations, homophones, and ellipses. Therefore, game intent recognition, as a key technology for converting player natural language into executable commands for the game system, directly determines the smoothness and immersion of game interaction through its recognition accuracy, response speed, and generalization ability.

[0003] The game intent recognition model trained using existing training schemes in related technologies cannot accurately capture the unique semantics of game language when performing intent recognition for game dialogues. The accuracy of intent recognition is low, and the real-time performance of intent recognition cannot be guaranteed. Summary of the Invention

[0004] This application provides a training method, apparatus, device, and storage medium for a game intent recognition model. It addresses the problems of existing game intent recognition model training schemes failing to accurately capture the unique semantics of game language, resulting in low intent recognition accuracy and inability to guarantee real-time intent recognition. The new method introduces the knowledge distillation function of a large language model, pre-divides different intent tasks, refines the granularity of intent recognition, improves the accuracy of game intent recognition, and balances accuracy and speed. It also introduces multiple instruction templates for the same intent task to enhance the model's generalization ability.

[0005] In a first aspect, embodiments of this application provide a method for training a game intent recognition model, the method comprising: Obtain a basic sample set, and call a pre-trained large language model to perform deep semantic parsing on the data in each basic sample in the basic sample set to obtain the full intent annotation information corresponding to each basic sample. Multiple instruction templates corresponding to different pre-generated intent tasks are concatenated with each of the base samples to obtain multiple input samples corresponding to each intent task. The full intent annotation information corresponding to each base sample is decomposed to obtain the intent annotation information corresponding to each input sample. Based on each input sample and the intent annotation information corresponding to each input sample, the pre-trained lightweight language model is subjected to supervised fine-tuning training for different intent tasks to obtain a trained game intent recognition model, which is used to output the intent recognition results of game dialogue.

[0006] Secondly, embodiments of this application also provide a training apparatus for a game intent recognition model, comprising: The acquisition module is configured to acquire the basic sample set; The parsing module is configured to call a pre-trained large language model to perform deep semantic parsing on the data in each basic sample in the basic sample set, and obtain the full intent annotation information corresponding to each basic sample. The splicing module is configured to splice multiple instruction templates corresponding to different pre-generated intent tasks with each of the basic samples to obtain multiple input samples corresponding to each intent task; The decomposition module is configured to decompose the full intent annotation information corresponding to each of the basic samples to obtain the intent annotation information corresponding to each of the input samples; The training module is configured to perform supervised fine-tuning training of a pre-trained lightweight language model for different intent tasks based on each input sample and the intent annotation information corresponding to each input sample, so as to obtain a trained game intent recognition model. The game intent recognition model is used to output the intent recognition results of game dialogue.

[0007] Thirdly, embodiments of this application also provide an electronic device, the device comprising: One or more processors; Storage device, configured to store one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the training method for the game intent recognition model described in the embodiments of this application.

[0008] Fourthly, embodiments of this application also provide a non-volatile storage medium for storing computer-executable instructions, which, when executed by a computer processor, are configured to perform the training method for the game intent recognition model described in embodiments of this application.

[0009] In this embodiment, a basic sample set is obtained, and a pre-trained large language model is invoked to perform deep semantic parsing on the data in each basic sample in the basic sample set to obtain the full intent annotation information corresponding to each basic sample. Multiple instruction templates corresponding to different pre-generated intent tasks are concatenated with each basic sample to obtain multiple input samples corresponding to each intent task. The full intent annotation information corresponding to each basic sample is then decomposed to obtain the intent annotation information corresponding to each input sample. Based on each input sample and the intent annotation information corresponding to each input sample, a pre-trained lightweight language model is subjected to supervised fine-tuning training for different intent tasks to obtain a trained game intent recognition model. The game intent recognition model is used to output the intent recognition results of game dialogues. This scheme can introduce the knowledge distillation effect of a large language model, pre-divide different intent tasks, refine the granularity of intent recognition, improve the accuracy of the model's game intent recognition, balance accuracy and speed, and simultaneously introduce multiple instruction templates for the same intent task to enhance the model's generalization ability. Attached Figure Description

[0010] Figure 1 A flowchart illustrating a training method for a game intent recognition model provided in this application embodiment; Figure 2 A flowchart illustrating a specific implementation process for obtaining full intent annotation information by annotating basic samples, as provided in this application embodiment; Figure 3 A flowchart illustrating a specific implementation process for determining a set of candidate terms corresponding to a basic sample and the target semantics corresponding to each candidate term, provided in this application embodiment; Figure 4 A flowchart illustrating a specific implementation process for determining a set of candidate terms corresponding to a basic sample and the target semantics corresponding to each candidate term based on a preset game terminology dictionary, provided in this application embodiment; Figure 5 A flowchart illustrating a specific implementation process for determining the statement dependency chain information corresponding to each basic sample, provided in this application embodiment; Figure 6 A flowchart illustrating a specific implementation process for obtaining intent structure information corresponding to each basic sample by decomposing the statement to be analyzed into intent structure, as provided in this application embodiment; Figure 7 A flowchart illustrating the specific implementation process of supervised fine-tuning training of a lightweight language model for different intent tasks, as provided in this application embodiment; Figure 8 This is a schematic diagram illustrating the model training method based on the game intent recognition model provided in the embodiments of this application. Figure 9A structural block diagram of a training device for a game intent recognition model provided in an embodiment of this application; Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0011] The embodiments of this application will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the embodiments of this application and are not intended to limit the scope of the embodiments. Furthermore, it should be noted that, for ease of description, only the parts relevant to the embodiments of this application are shown in the accompanying drawings, not the entire structure.

[0012] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0013] The training method for the game intent recognition model provided in this application embodiment can be executed by a computer device. The computer device refers to any electronic device with data computing, processing and storage capabilities, such as a server. This application embodiment does not limit this.

[0014] Figure 1 A flowchart illustrating a training method for a game intent recognition model provided in this application embodiment is shown below. Figure 1 As shown, the training method for this game intent recognition model specifically includes the following steps: Step S101: Obtain the basic sample set, call the pre-trained large language model to perform deep semantic parsing on the data in each basic sample in the basic sample set, and obtain the full intent annotation information corresponding to each basic sample.

[0015] The basic sample set can include multiple basic samples, each of which can include historical dialogue sequences, statements to be analyzed, and game metadata. The historical dialogue sequence can be multiple player dialogue records preceding the statement to be analyzed, reflecting the context of the dialogue. The statement to be analyzed can be the current player's statement that requires intent recognition. Game metadata can be structured information related to the game scene, including game type, game mode, battle stage, player role, etc. The pre-trained large language model can be a language model with hundreds of billions of parameters (such as the DeepSeek series, GPT series, etc.) pre-trained on a large-scale general text corpus, possessing powerful general semantic understanding and reasoning capabilities. By utilizing the pre-trained large language model, implicit semantics, contextual relationships, and logical connections can be mined based on the recognition of text content, allowing for structured annotation of all dimensions of intent information in the game dialogue, resulting in full-scale intent annotation information. Full-scale intent annotation information can be labeled data covering different aspects such as terminology semantics, contextual dependencies, and multi-intent distribution.

[0016] In one embodiment, a specific implementation process for annotating basic samples to obtain full intent annotation information is described. Please refer to [reference needed]. Figure 2 This is a flowchart illustrating a specific implementation process for obtaining full intent annotation information by annotating basic samples, as provided in an embodiment of this application. Figure 2 As shown, the basic samples include historical dialogue sequences, statements to be analyzed, and game metadata. The specific implementation process of calling a pre-trained large language model to perform deep semantic parsing on the data in each basic sample set to obtain the full intent annotation information corresponding to each basic sample includes, but is not limited to, the following steps S201-S203. It should be noted that steps S201-S203 are all executed by the large language model, specifically by instructing the large language model to process according to the steps using preset prompt words. Step S201: Based on the game metadata in each basic sample, extract terms from the historical dialogue sequence and the statement to be analyzed to obtain the candidate term set corresponding to each basic sample and the target semantics corresponding to each candidate term in the candidate term set.

[0017] Introducing game metadata provides crucial contextual information, helping the model distinguish the different semantics of the same words in different game scenarios and resolving semantic ambiguity in game language. Term extraction involves identifying words and phrases with specific game meanings from game dialogue. The candidate term set is a collection of game terms and phrases. The target semantics is the precise semantic meaning of the candidate terms within the current game scenario.

[0018] In one embodiment, a specific implementation process for determining the candidate term set corresponding to the base sample and the target semantics corresponding to each candidate term is described. Please refer to [reference needed]. Figure 3 This is a flowchart illustrating a specific implementation process for determining the set of candidate terms corresponding to a basic sample and the target semantics corresponding to each candidate term, as provided in this application embodiment. Figure 3 As shown, based on the game metadata in each basic sample, term extraction is performed on the historical dialogue sequence and the statement to be analyzed to obtain the candidate term set corresponding to each basic sample and the target semantics corresponding to each candidate term in the candidate term set. The specific implementation process includes, but is not limited to, the following steps S301-S303: Step S301: Based on the preset game terminology dictionary, extract terms from the historical dialogue sequence and the statement to be analyzed corresponding to each basic sample to obtain a candidate term set corresponding to each basic sample and at least one semantic meaning corresponding to each candidate term in the candidate term set.

[0019] The preset game terminology dictionary can be a pre-collected and organized dictionary containing various game terms (such as game slang, abbreviations, and homophones) and their semantic explanations, covering commonly used terms in mainstream game genres. In one embodiment, based on the words and phrases in the preset game terminology dictionary, the existence of words or phrases recorded in the preset game terminology dictionary can be searched in the historical dialogue sequence and the statement to be analyzed corresponding to each basic sample, and these words or phrases can be extracted as a candidate term set. Each candidate term can be queried in the preset game terminology dictionary to find multiple corresponding semantics. It should be noted that the same term may express different meanings for different game genres, game modes, etc.; therefore, the same term may correspond to one or more semantics.

[0020] In one embodiment, terms recorded in a preset game terminology dictionary are extracted from the historical dialogue sequence and the statement to be analyzed corresponding to each basic sample, resulting in a candidate term set corresponding to each basic sample and at least one semantic meaning corresponding to each candidate term in the candidate term set.

[0021] In one embodiment, a specific implementation process is described for determining the candidate term set corresponding to the basic sample and the target semantics corresponding to each candidate term based on a preset game terminology dictionary. Please refer to [reference needed]. Figure 4 This is a flowchart illustrating a specific implementation process for determining a set of candidate terms corresponding to a basic sample and the target semantics corresponding to each candidate term based on a preset game terminology dictionary, as provided in this application embodiment. Figure 4As shown, the specific implementation process of extracting terms from the historical dialogue sequence and the statement to be analyzed corresponding to each basic sample based on a preset game terminology dictionary, to obtain a candidate term set for each basic sample and at least one semantic meaning corresponding to each candidate term in the candidate term set, includes but is not limited to the following steps S401-S403: Step S401: Extract the first term set recorded in the preset game term dictionary and at least one semantic meaning corresponding to each term in the first term set from the historical dialogue sequence and the statement to be analyzed corresponding to each basic sample.

[0022] The first terminology set can be a set of terms extracted from the dialogue and existing in a preset game terminology dictionary. By extracting terms from the preset game terminology dictionary, most known game terms can be quickly covered.

[0023] Step S402: Perform high-frequency vocabulary statistics on the historical dialogue sequence and the statement to be analyzed corresponding to each basic sample to obtain a second term set, and determine at least one semantic meaning corresponding to each term in the second term set through the set network search path.

[0024] The high-frequency vocabulary statistics can be achieved by statistically analyzing the most frequently occurring words and phrases in the dialogue text, and then filtering these words and phrases in conjunction with known non-technical terms or phrases to retain unique words or phrases not included in the preset game terminology dictionary. These unique words or phrases can be newly emerging game terms. The second terminology set can be a newly added set of terms that are not present in the preset game terminology dictionary. The web search path can be a web resource path used to query the semantics of words or phrases, such as search engines, game forums, etc., to supplement the latest terminology meanings and compensate for the lag in updating the preset game terminology dictionary.

[0025] Step S403: Merge the first term set and the second term set corresponding to each basic sample to obtain the candidate term set corresponding to each basic sample and at least one semantic meaning corresponding to each candidate term in the candidate term set.

[0026] Combining the first and second term sets yields a complete candidate term set. Each candidate term in the candidate term set matches at least one corresponding semantic meaning.

[0027] As can be seen from steps S401-S403 above, by using a preset game terminology dictionary to extract the terms contained in the historical dialogue sequence and the statement to be analyzed, most of the known game terms can be quickly covered; by timely identifying game terms in the historical dialogue sequence and the statement to be analyzed that are not recorded in the preset game terminology dictionary, the timeliness of the data can be guaranteed, and the game terms that may appear in the historical dialogue sequence and the statement to be analyzed can be more comprehensively covered; by merging the first terminology set and the second terminology set, the identification of known terms and new terms can be taken into account, providing a complete foundation for subsequent semantic analysis.

[0028] Step S302: Based on the game type and game mode in the game metadata corresponding to each basic sample, filter at least one semantic corresponding to each candidate term to obtain the candidate semantic corresponding to each candidate term.

[0029] The game type can correspond to a specific game category, such as role-playing games or multiplayer tactical battle royale games. The game mode can be a specific gameplay mode within a game, such as 5v5 ranked mode or 3v3 matchmaking mode. For example, regarding the game term "point," in game A, it can mean a location where a specific action needs to be performed, while in game B, it can mean a resource-rich area or terrain on the map. As another example, regarding the game term "point," in game mode A1 of game A, it can mean a location where a specific item needs to be placed, while in game mode B1 of game B, it can mean a location where a specific target needs to be defended. Therefore, based on the game type and game mode in the game metadata, semantic filtering can be performed on each candidate term, retaining candidate semantics that match the player's current game state.

[0030] Step S303: Extract the context information of the statement to be analyzed from the historical dialogue sequence corresponding to each basic sample, and perform semantic supplementation on the candidate semantics corresponding to each candidate term based on the context information to obtain the target semantics corresponding to each candidate term corresponding to each basic sample.

[0031] The contextual information can be several dialogue records before and after the statement to be analyzed, providing more specific contextual information. Semantic supplementation can further clarify the semantics of terms using contextual information. It should be noted that contextual information further refines the semantics of candidate terms by providing the following key clues: subject and object of the action (who is performing the action, and to whom is the action directed), time and state (current combat phase, skill cooldown status, character health status), cause and effect (what events happened before, and what goal the player hopes to achieve), logical connection (causal, conditional, and progressive relationships between the current statement and historical statements), etc. For example, for the game term "smoke screen," its candidate semantics are determined to be "throwing a smoke grenade." For example, the contextual information is: Player A ("I rushed too low"), Player B ("smoke screen"). According to this contextual information, Player A is launching an attack and needs teammates to provide support. Therefore, the candidate semantics can be further adjusted to "throwing a smoke grenade to teammates to cover the attack." For example, the context information is: Player A ("The bomb has been planted, defend the bomb"), Player B ("Smoke bomb"). According to this context information, Player A has successfully planted the bomb and needs to defend it. Therefore, the candidate semantics can be further adjusted to "Throw smoke bombs to block the bomb site and prevent the enemy from defusing the bomb".

[0032] As can be seen from steps S301-S303 above, by using a preset game terminology dictionary for term extraction, known terms can be quickly identified and their semantics extracted and obtained; by filtering out semantics that are irrelevant to the current game type and mode, the candidate semantics that best match the current game state can be determined from at least one semantic; by using contextual information for semantic supplementation, the special usage and extended meaning of terms in a specific context can be accurately identified, thereby improving the accuracy of semantic understanding.

[0033] Step S202: Perform context association parsing on the historical dialogue sequence and the statement to be analyzed corresponding to each basic sample to obtain the statement dependency chain information corresponding to each basic sample.

[0034] Contextual analysis can be used to analyze the semantic and logical relationships between historical dialogues and the current statement. Statement dependency chain information can be structured information reflecting the dependencies between historical and current statements, including related statements and logical relationship types.

[0035] In one embodiment, a specific implementation process for determining the statement dependency chain information corresponding to each basic sample is described. Please refer to [reference needed]. Figure 5 This is a flowchart illustrating a specific implementation process for determining the statement dependency chain information corresponding to each basic sample, as provided in an embodiment of this application. Figure 5As shown, the specific implementation process of performing context association parsing on the historical dialogue sequence and the statement to be analyzed corresponding to each basic sample to obtain the statement dependency chain information corresponding to each basic sample includes, but is not limited to, the following steps S501-S503: Step S501: Perform semantic similarity matching between each historical speech statement in the historical dialogue sequence corresponding to each basic sample and the statement to be analyzed, to obtain the set of associated statements corresponding to each basic sample.

[0036] Semantic similarity matching can be a process of calculating the semantic similarity between two statements. The higher the similarity value, the stronger the semantic connection between the two statements. Specific methods such as vector similarity calculation and keyword matching can be used, which are not limited in this application. The set of related statements can be a set of statements in the historical dialogue sequence that have a high semantic similarity to the statement to be analyzed.

[0037] Step S502: Identify the logical relationship between the set of associated statements corresponding to each basic sample and the statement to be analyzed, and obtain the statement logical relationship information corresponding to each basic sample.

[0038] The logical relationships can include causal relationships, conditional relationships, progressive relationships, and adversative relationships. In one embodiment, a pre-trained relationship classification model is used to identify the logical relationship type between the set of associated statements corresponding to each basic sample and the statement to be analyzed, thereby obtaining the logical relationship information of the statements. In another embodiment, conjunctions (such as "because," "therefore," "however," etc.) are extracted from the statements, and the logical relationships are inferred based on the specific meanings of the conjunctions.

[0039] Step S503: Based on the statement logical relationship information corresponding to each basic sample, combine the associated statement set and the statement to be analyzed to obtain statement dependency chain information.

[0040] The statement dependency chain information can be a chain structure that organizes related statements and statements to be analyzed according to logical relationships, reflecting the semantic flow and dependencies of the dialogue. Specifically, using the logical relationship information of statements as the basis for combination, each statement in the set of related statements is split and rearranged and combined with the statement to be analyzed to obtain the statement dependency chain information. For example, statement A — (progressive) — statement B — (conditional) statement C.

[0041] As can be seen from steps S501-S503 above, by filtering related statements through semantic similarity matching, historical dialogues that are semantically irrelevant to the current statement can be filtered out, focusing on relevant contextual information; by identifying the logical relationship between statements, the dialogue logic can be understood in depth, and the logical relationship implied between statements can be clarified; by constructing statement dependency chains, a clear chain structure can be provided for subsequent intent decomposition, improving the accuracy of subsequent intent recognition.

[0042] Step S203: Based on the target semantics and statement dependency chain information corresponding to each candidate term of each basic sample, perform intent decomposition on the statement to be analyzed to obtain the intent structure information corresponding to each basic sample.

[0043] Intent decomposition can be the process of breaking down complex player intentions into multiple sub-intentions or intentions of different dimensions. Intent structure information can be a structured representation of intentions that includes information such as intention graphs, sub-intentions, and intention probability distributions.

[0044] As can be seen from the above steps S201-S203, by combining game metadata for term extraction, irrelevant words can be filtered out, and game terms and related semantics can be fully covered; by performing context association parsing, the dependency relationship between various statements in the dialogue can be analyzed, and the dialogue logic chain can be provided; by performing intent decomposition, the multi-intent problem in the statement to be analyzed can be handled, and more accurate and reliable intent information can be provided.

[0045] In one embodiment, a specific implementation process is described for decomposing the statement to be analyzed to obtain the intent structure information corresponding to each basic sample. Please refer to [reference needed]. Figure 6 This is a flowchart illustrating a specific implementation process for decomposing a statement to be analyzed into intent structure information corresponding to each basic sample, as provided in an embodiment of this application. Figure 6 As shown, the specific implementation process of obtaining the intent structure information corresponding to each basic sample by performing intent decomposition on the statement to be analyzed based on the target semantics and statement dependency chain information corresponding to each candidate term of each basic sample includes, but is not limited to, the following steps S601-S603: Step S601: Based on the target semantics and statement dependency chain information corresponding to each candidate term of each basic sample, generate a candidate intent list corresponding to each basic sample.

[0046] Specifically, the process of generating a candidate intent list for each basic sample can be as follows: The target semantics of all candidate terms are structurally decomposed to extract four core elements: action, entity, attribute, and role, forming a semantic element set. Using the statement dependency chain information as a framework, the extracted semantic element set is filled into the text positions adjacent to the original terms to obtain fused statement semantic information. The fused statement semantic information is then matched and reasoned with a preset game intent template to generate a candidate intent list. The preset game intent template covers all common intent types in the game, such as: tactical commands (initiate attack, request support, retreat, assemble); status reports (self-status report, enemy status report, map information report); resource requests (request buffs, request healing, request equipment); and communication / discussion (tactical discussion, decision suggestions, emotional expression).

[0047] Step S602: Perform intent distribution prediction on the candidate intent list corresponding to each basic sample to obtain the intent probability distribution corresponding to each basic sample.

[0048] The intent distribution prediction process involves quantifying and scoring each candidate intent in the candidate intent list using term frequency features and contextual information content features, and then converting the scores into a standardized probability distribution. Specifically, term frequency features measure the strength of the association between a candidate intent and game terms appearing in the current dialogue; the higher the frequency of highly relevant terms in the dialogue, the greater the likelihood of that intent. Contextual information content measures the degree of information support obtained from the statement dependency chain; the more nodes in the statement dependency chain the intent is semantically related to, the greater the information content and the higher the likelihood. For term frequency features, the frequency of at least one associated term for each candidate intent is calculated relative to all terms in the historical dialogue sequence and the statement to be analyzed, and the frequency result is used as the term frequency feature score for that candidate intent. For contextual information content features, the percentage of all contextual information associated with each candidate intent relative to the overall information content of the historical dialogue sequence and the statement to be analyzed is calculated, and the percentage result is used as the contextual information content feature score for that candidate intent. Then, the term frequency feature score and the contextual information content feature score are weighted and calculated to obtain the target score corresponding to the candidate intent. Finally, the target scores corresponding to each candidate intent in the candidate intent list are uniformly normalized to obtain the intent probability distribution, which records the probability corresponding to each candidate intent.

[0049] Step S603: Based on the intent probability distribution corresponding to each basic sample, the semantic unit decomposition of the statement to be analyzed is performed to obtain the intent structure information corresponding to each basic sample.

[0050] Semantic units can be the smallest linguistic fragments carrying independent game semantic information, specifically single words, phrases, game slang or abbreviations, ellipses, etc. Optionally, semantic units can be divided into action units, entity units, attribute units, role units, condition units, and modifier units. For example, action units can represent game actions performed or requested by the player, such as "attack," "push," "defend," etc.; entity units can represent specific objects in the game, such as "tower," "jungle," etc.; attribute units can represent the state or characteristics of an entity, such as "refreshed," "ultimate skill depleted," etc.; role units can represent the executor or recipient of an action, such as "teammate," "we," etc.; condition units can represent the preconditions for an action, such as "if," "unless," etc.; and modifier units can represent the degree, time, or manner of an action, such as "quickly," "immediately," etc. Specifically, the statement to be analyzed can undergo basic word segmentation, and the segmentation results can be semantically identified and type-labeled to obtain the various semantic units corresponding to the statement to be analyzed. Each semantic unit can be matched with each candidate intent in the intent probability distribution to form a semantic unit-candidate intent correspondence and obtain intent structure information.

[0051] As can be seen from the above steps S601-S603, by generating a candidate intent list corresponding to the basic samples, all possible intents can be fully covered, providing a basis for subsequent probability prediction; by predicting the intent distribution, the confidence of candidate intents can be quantified; by performing semantic unit decomposition and generating intent structure information, the accurate distribution of possible intents in the statement to be analyzed can be clearly defined.

[0052] refer to Figure 1 It also includes step S102, which involves concatenating multiple instruction templates corresponding to different pre-generated intent tasks with each basic sample to obtain multiple input samples corresponding to each intent task, and decomposing the full intent annotation information corresponding to each basic sample to obtain the intent annotation information corresponding to each input sample.

[0053] The lightweight language model can employ a multi-task learning framework, which includes a main task (intent recognition) and multiple auxiliary tasks (such as slang translation, emotion recognition, action suggestion, role inference, and timeliness judgment). The slang translation task might require the model to translate game terminology into standard semantic expressions; the emotion recognition task might analyze the emotional tendency and intensity of player dialogue; the action suggestion task might derive recommended operation commands based on the currently recognized intent; the role inference task might identify the player's role in the game (e.g., commander, damage dealer, support); and the timeliness judgment task might distinguish between intents requiring immediate response (e.g., calls for help) and non-immediate intents (e.g., tactical discussions). For each intent task, one or more instruction templates can be set. By employing differentiated sentence structures, terminology, and task description perspectives, the model can robustly understand various instruction variations. The instruction templates and basic samples are concatenated to form input samples, which can be used as input data for model training. Specifically, the input samples can consist of "instruction template + historical dialogue sequence + statement to be analyzed + game metadata". Since different input samples correspond to different intent tasks, it is possible to extract annotation data for a single task dimension corresponding to a specific input sample from the full set of intent annotation information.

[0054] Step S103: Based on each input sample and the intent annotation information corresponding to each input sample, perform supervised fine-tuning training on the pre-trained lightweight language model for different intent tasks to obtain the trained game intent recognition model, wherein the game intent recognition model is used to output the intent recognition results of game dialogue.

[0055] The lightweight language model can be a pre-trained language model with a small number of parameters (e.g., less than 7 bytes). All intent tasks within this lightweight language model can share the same backbone model architecture, but each intent task can have its own independent output layer. Optionally, supervised fine-tuning training can employ a low-rank adaptation method to achieve efficient parameter updates. Specifically, an adapter module is added in parallel as a bypass to each linear transformation layer of the lightweight language model (including Query / Key / Value / Output projections in the self-attention layer and up / down projections in the FFN layer). This adapter module can be set to a rank parameter of 32 and a scaling factor of 64, while freezing the original weights obtained from pre-training of the lightweight language model and training only the newly added adapter parameters and the output layer weights specific to each intent task. The optimization objective can be a weighted sum of the cross-entropy losses of each intent task, with higher weights assigned to the main intent task and relatively lower weights to the auxiliary intent tasks. The Adam optimizer is used with multiple iterations at a low learning rate (e.g., 5e-5) to ultimately obtain a lightweight language model with powerful intent recognition and semantic understanding capabilities.

[0056] In one embodiment, a specific implementation process for supervised fine-tuning training of a lightweight language model for different intent tasks is described. Please refer to [reference needed]. Figure 7 This is a flowchart illustrating the specific implementation process of supervised fine-tuning training of a lightweight language model for different intent tasks, as provided in this application embodiment. Figure 7 As shown, the specific implementation process of training a pre-trained lightweight language model for different intent tasks based on each input sample and the intent annotation information corresponding to each input sample, to obtain the trained game intent recognition model, includes but is not limited to the following steps S701-S703: Step S701: Sample multiple input samples according to the task sampling rate corresponding to different intention tasks to obtain multiple target input samples for the current training batch.

[0057] The task sampling rate represents the probability of each intent task being sampled in a training batch, reflecting the importance of that task during training. A training batch can be the set of samples used by the lightweight language model for each parameter update. The target input sample can be a sample obtained from all input samples for training the current batch. Specifically, in the early stages of training, relatively simple auxiliary intent tasks (such as emotion recognition and slang translation) are sampled first to guide the model to adapt to the multi-task framework with a lower learning threshold. Then, as training rounds progress, the sampling probability of difficult tasks (such as intent recognition and action suggestion) is gradually increased to guide the model to master complex capabilities step by step.

[0058] Step S702: Input multiple target input samples into a pre-trained lightweight language model to obtain the intent prediction result corresponding to each target input sample.

[0059] The intent prediction result can be the result of a lightweight language model predicting the intent of an input sample, including information such as intent type and probability.

[0060] Step S703: Based on the intent prediction results and corresponding intent annotation information of each target input sample, calculate the loss value corresponding to different intent tasks, and update the parameters of the lightweight language model based on the loss value corresponding to different intent tasks until the loss converges to obtain the trained game intent recognition model.

[0061] In each training round, the loss of each target input sample is calculated based on the intent prediction result and intent annotation information to obtain the loss value of that target input sample. For each intent task, the loss values ​​calculated from one or more corresponding target input samples are averaged to obtain the mean loss for each intent task. Then, the mean losses for each intent task are weighted to obtain the target loss value. The parameters of the lightweight language model are updated based on the gradient of this target loss value until the loss converges after multiple training batches, resulting in the trained game intent recognition model.

[0062] As can be seen from steps S701-S703 above, by using the task sampling rate to sample samples, the sampling rate can be adjusted according to the importance and difficulty of the task, ensuring that important and difficult tasks are trained sufficiently. Furthermore, progressive training can ensure that the model can better cope with various intention tasks. By inputting samples into the model to obtain prediction results and performing loss calculations and parameter updates for different intention tasks, multi-task joint optimization can be achieved, thereby improving model performance.

[0063] Optionally, during implementation, player dialogue flow can be monitored in real time. By combining word frequency statistics algorithms (for identifying high-frequency new words), information entropy calculation (for evaluating the information value of new words in context), and context consistency detection (for confirming the stable occurrence of new words in similar contexts), out-of-vocabulary terms (i.e., new game slang or terminology not learned by the game intent recognition model) appearing in the dialogue can be automatically identified. The calculation results obtained from the aforementioned word frequency statistics algorithm, information entropy calculation, and context consistency detection are fused to obtain the detection confidence. When the detection confidence exceeds a preset threshold, a lightweight model fine-tuning process is initiated. Specifically, a few-shot learning strategy is adopted, using manual or semi-automatic annotation of 50-100 typical dialogue samples containing the new term, and a low-rank adaptation method is used to quickly and incrementally update the game intent recognition model to efficiently integrate the new terminology knowledge into the existing model. To ensure system stability, the current stable version can be automatically backed up before the model is updated. During the update process, strict verification is performed. If the model's performance on the validation set deteriorates significantly or anomalies after the update, the system will automatically roll back to the stable version before the update, ensuring the reliability and continuity of online services.

[0064] As can be seen from steps S101-S103 above, by using a large language model for deep semantic parsing, the annotation cost can be reduced and the annotation quality can be improved; by concatenating instruction templates and decomposing annotations, the model can understand the same task expressed in different ways, enhance generalization ability, and also support the model's multi-task learning; by performing multi-task supervised fine-tuning on the lightweight language model, the lightweight speech model can have intent recognition ability, maintaining high accuracy while improving inference speed, and achieving a balance between accuracy and speed.

[0065] Figure 8 This is a schematic diagram illustrating the model training method based on the game intent recognition model provided in the embodiments of this application. Figure 8 As shown, the pre-trained large language model 801 is called to perform deep semantic parsing on the basic sample set 802 to obtain the full intent annotation information 803 corresponding to the basic sample set 802. Multiple instruction templates 804 corresponding to different intent tasks are concatenated with the basic samples in the basic sample set 802 to obtain the input sample set 805. The full intent annotation information 803 is decomposed to obtain the intent annotation information 806 corresponding to each input sample in the input sample set 805. Based on the input sample set 805 and the intent annotation information 806 corresponding to each input sample, the pre-trained lightweight language model 807 is subjected to supervised fine-tuning training for different intent tasks to obtain the trained game intent recognition model 808.

[0066] Figure 9 This is a structural block diagram of a training device for a game intent recognition model provided in an embodiment of this application. The device is configured to execute the training method for the game intent recognition model provided in the above embodiment, and has corresponding functional modules and beneficial effects for executing the method. For example... Figure 9 As shown, the device specifically includes: Module 901 is configured to acquire the basic sample set; The parsing module 902 is configured to call a pre-trained large language model to perform deep semantic parsing on the data in each basic sample in the basic sample set, and obtain the full intent annotation information corresponding to each basic sample. The splicing module 903 is configured to splice multiple instruction templates corresponding to different pre-generated intent tasks with each basic sample to obtain multiple input samples corresponding to each intent task; The decomposition module 904 is configured to decompose the full intent annotation information corresponding to each basic sample to obtain the intent annotation information corresponding to each input sample; Training module 905 is configured to perform supervised fine-tuning training of a pre-trained lightweight language model for different intent tasks based on each input sample and the intent annotation information corresponding to each input sample, so as to obtain a trained game intent recognition model. The game intent recognition model is used to output the intent recognition results of game dialogue.

[0067] As can be seen from steps S101-S103 above, by using a large language model for deep semantic parsing, the annotation cost can be reduced and the annotation quality can be improved; by concatenating instruction templates and decomposing annotations, the model can understand the same task expressed in different ways, enhance generalization ability, and also support the model's multi-task learning; by performing multi-task supervised fine-tuning on the lightweight language model, the lightweight speech model can have intent recognition ability, maintaining high accuracy while improving inference speed, and achieving a balance between accuracy and speed.

[0068] In one possible embodiment, the base sample includes historical dialogue sequences, statements to be analyzed, and game metadata; the parsing module 902 is specifically configured as follows: The pre-trained large language model is invoked to perform deep semantic parsing on the data in each basic sample in the basic sample set, obtaining the full intent annotation information corresponding to each basic sample, including: Based on the game metadata in each base sample, term extraction is performed on the historical dialogue sequence and the statement to be analyzed to obtain the candidate term set corresponding to each base sample and the target semantics corresponding to each candidate term in the candidate term set. For each basic sample, the historical dialogue sequence and the statement to be analyzed are parsed to obtain the statement dependency chain information for each basic sample. Based on the target semantics and statement dependency chain information corresponding to each candidate term of each base sample, the statement to be analyzed is decomposed into intent structure information corresponding to each base sample.

[0069] In one possible embodiment, the parsing module 902 is further configured as follows: Based on a pre-defined game terminology dictionary, terminology is extracted from the historical dialogue sequence and the statement to be analyzed corresponding to each basic sample, resulting in a candidate terminology set for each basic sample and at least one semantic meaning corresponding to each candidate term in the candidate terminology set. Based on the game type and game mode in the game metadata corresponding to each basic sample, at least one semantics corresponding to each candidate term is filtered to obtain the candidate semantics corresponding to each candidate term. The context information of the statement to be analyzed is extracted from the historical dialogue sequence corresponding to each basic sample, and the candidate semantics corresponding to each candidate term is semantically supplemented based on the context information to obtain the target semantics corresponding to each candidate term for each basic sample.

[0070] In one possible embodiment, the parsing module 902 is further configured as follows: Extract the first term set recorded in the preset game term dictionary and at least one semantic meaning corresponding to each term in the first term set from the historical dialogue sequence and the statement to be analyzed corresponding to each basic sample; A second term set is obtained by performing high-frequency vocabulary statistics on the historical dialogue sequence and the statement to be analyzed corresponding to each basic sample, and at least one semantic meaning corresponding to each term in the second term set is determined by the set network search path. The first term set and the second term set corresponding to each basic sample are merged to obtain the candidate term set corresponding to each basic sample and at least one semantic meaning corresponding to each candidate term in the candidate term set.

[0071] In one possible embodiment, the parsing module 902 is further configured as follows: Each historical speech statement in the historical dialogue sequence corresponding to each basic sample is semantically similarly matched with the statement to be analyzed to obtain the set of associated statements corresponding to each basic sample. Logical relationship identification is performed between the set of associated statements corresponding to each basic sample and the statement to be analyzed, thus obtaining the statement logical relationship information corresponding to each basic sample; Based on the logical relationship information of the statements corresponding to each basic sample, the set of related statements and the statements to be analyzed are combined to obtain the statement dependency chain information.

[0072] In one possible embodiment, the parsing module 902 is further configured as follows: Based on the target semantics and statement dependency chain information of each candidate term corresponding to each base sample, a candidate intent list corresponding to each base sample is generated. For each base sample, predict the intent distribution of the candidate intent list to obtain the intent probability distribution for each base sample. Based on the intent probability distribution corresponding to each basic sample, the semantic unit decomposition of the statement to be analyzed is performed to obtain the intent structure information corresponding to each basic sample.

[0073] In one possible embodiment, the training module 905 is further configured as follows: Multiple input samples are sampled according to the task sampling rate corresponding to different intention tasks to obtain multiple target input samples for the current training batch; Multiple target input samples are input into a pre-trained lightweight language model to obtain the intent prediction result for each target input sample; Based on the intent prediction results and corresponding intent annotation information of each target input sample, loss is calculated to obtain the loss value corresponding to different intent tasks. Then, the parameters of the lightweight language model are updated based on the loss value corresponding to different intent tasks until the loss converges to obtain the trained game intent recognition model.

[0074] Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application, such as... Figure 10 As shown, the device includes a processor 1001, a memory 1002, an input device 1003, and an output device 1004; the number of processors 1001 in the device can be one or more. Figure 10 Taking a processor 1001 as an example; the processor 1001, memory 1002, input device 1003, and output device 1004 in the device can be connected via a bus or other means. Figure 10 Taking a bus connection as an example, the memory 1002, as a computer-readable storage medium, can be configured to store software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the training method of the game intent recognition model in this embodiment. The processor 1001 executes various functional applications and data processing of the device by running the software programs, instructions, and modules stored in the memory 1002, thereby implementing the above-mentioned training method of the game intent recognition model. The input device 1003 can be configured to receive input digital or character information and generate key signal inputs related to the user settings and function control of the device. The output device 1004 may include a display screen or other display device.

[0075] The electronic device provided above can be used to execute the training method of the game intent recognition model provided in any of the above embodiments, and has corresponding functions and beneficial effects.

[0076] This application also provides a non-volatile storage medium containing computer-executable instructions. When executed by a computer processor, the computer-executable instructions are configured to execute a training method for a game intent recognition model described in the above embodiments. The method includes: acquiring a basic sample set; calling a pre-trained large language model to perform deep semantic parsing on the data in each basic sample in the basic sample set to obtain full intent annotation information corresponding to each basic sample; concatenating multiple pre-generated instruction templates corresponding to different intent tasks with each basic sample to obtain multiple input samples corresponding to each intent task; decomposing the full intent annotation information corresponding to each basic sample to obtain intent annotation information corresponding to each input sample; and performing supervised fine-tuning training of the pre-trained lightweight language model for different intent tasks based on each input sample and the intent annotation information corresponding to each input sample to obtain a trained game intent recognition model. The game intent recognition model is used to output the intent recognition result of game dialogue.

[0077] Storage medium – any type of memory device or storage device. The term “storage medium” is intended to include: mounting media, such as CD-ROMs, floppy disks, or magnetic tape devices; computer system memory or random access memory, such as DRAM, DDR RAM, SRAM, EDO RAM, Rambus RAM, etc.; non-volatile memory, such as flash memory, magnetic media, optical storage; registers or other similar types of memory elements, etc. Storage media may also include other types of memory or combinations thereof. Furthermore, storage media may reside in a first computer system in which the program is executed, or may reside in a different second computer system connected to the first computer system via a network (such as the Internet). The second computer system can provide program instructions to the first computer for execution. The term “storage medium” can include two or more storage media residing in different locations (e.g., in different computer systems connected via a network). Storage media may store program instructions (e.g., specifically implemented as a computer program) executable by one or more processors.

[0078] Of course, the computer-executable instructions provided in the embodiments of this application are not limited to the training method of the game intent recognition model as described above, but can also execute related operations in the training method of the game intent recognition model provided in any embodiment of this application.

[0079] It should be noted that the numbering of each step in this solution is only used to describe the overall design framework of this solution and does not indicate a necessary sequential relationship between the steps. As long as the overall implementation process conforms to the overall design framework of this solution, it falls within the protection scope of this solution. The literal order in the description is not an exclusive limitation on the specific implementation process of this solution. Those skilled in the art should understand that the embodiments of this application can be provided as methods, systems, or computer program products. In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory. Memory may include non-persistent memory in computer-readable media, random access memory (RAM), and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0080] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0081] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A training method for a game intent recognition model, characterized in that, include: Obtain a basic sample set, and call a pre-trained large language model to perform deep semantic parsing on the data in each basic sample in the basic sample set to obtain the full intent annotation information corresponding to each basic sample. Multiple instruction templates corresponding to different pre-generated intent tasks are concatenated with each of the base samples to obtain multiple input samples corresponding to each intent task. The full intent annotation information corresponding to each base sample is decomposed to obtain the intent annotation information corresponding to each input sample. Based on each input sample and the intent annotation information corresponding to each input sample, the pre-trained lightweight language model is subjected to supervised fine-tuning training for different intent tasks to obtain a trained game intent recognition model, which is used to output the intent recognition results of game dialogue.

2. The training method for the game intent recognition model according to claim 1, characterized in that, The basic samples include historical dialogue sequences, statements to be analyzed, and game metadata; The pre-trained large language model is invoked to perform deep semantic parsing on the data in each basic sample in the basic sample set, obtaining the full intent annotation information corresponding to each basic sample, including: Based on the game metadata in each of the basic samples, term extraction is performed on the historical dialogue sequences and the statements to be analyzed to obtain a candidate term set corresponding to each of the basic samples and the target semantics corresponding to each candidate term in the candidate term set. Context association parsing is performed on the historical dialogue sequence and the statement to be analyzed corresponding to each basic sample to obtain the statement dependency chain information corresponding to each basic sample; Based on the target semantics and statement dependency chain information corresponding to each candidate term of each base sample, the statement to be analyzed is decomposed to obtain the intent structure information corresponding to each base sample.

3. The training method for the game intent recognition model according to claim 2, characterized in that, The step involves extracting terms from historical dialogue sequences and statements to be analyzed based on game metadata in each base sample, resulting in a candidate term set for each base sample and the target semantics corresponding to each candidate term in the candidate term set, including: Based on a preset game terminology dictionary, terminology is extracted from the historical dialogue sequence and the statement to be analyzed corresponding to each basic sample to obtain a candidate terminology set corresponding to each basic sample and at least one semantic meaning corresponding to each candidate term in the candidate terminology set. Based on the game type and game mode in the game metadata corresponding to each of the basic samples, at least one semantic corresponding to each of the candidate terms is filtered to obtain the candidate semantic corresponding to each of the candidate terms. Context information of the statement to be analyzed is extracted from the historical dialogue sequence corresponding to each of the basic samples, and the candidate semantics corresponding to each candidate term are semantically supplemented based on the context information to obtain the target semantics corresponding to each candidate term corresponding to each of the basic samples.

4. The training method for the game intent recognition model according to claim 3, characterized in that, The step involves extracting terms from the historical dialogue sequence and the statement to be analyzed corresponding to each basic sample based on a preset game terminology dictionary, to obtain a candidate term set corresponding to each basic sample and at least one semantic meaning corresponding to each candidate term in the candidate term set, including: Extract from the historical dialogue sequence and the statement to be analyzed corresponding to each of the basic samples a first term set recorded in the preset game term dictionary and at least one semantic meaning corresponding to each term in the first term set; A second term set is obtained by performing high-frequency vocabulary statistics on the historical dialogue sequence and the statement to be analyzed corresponding to each basic sample, and at least one semantic meaning corresponding to each term in the second term set is determined by the set network search path. The first term set and the second term set corresponding to each basic sample are merged to obtain a candidate term set corresponding to each basic sample and at least one semantic corresponding to each candidate term in the candidate term set.

5. The training method for the game intent recognition model according to claim 2, characterized in that, The step of performing context association parsing on the historical dialogue sequence and the statement to be analyzed corresponding to each basic sample to obtain the statement dependency chain information corresponding to each basic sample includes: Each historical speech statement in the historical dialogue sequence corresponding to each basic sample is semantically similarly matched with the statement to be analyzed to obtain the set of associated statements corresponding to each basic sample. Logical relationship identification is performed on the set of associated statements corresponding to each basic sample and the statement to be analyzed to obtain the statement logical relationship information corresponding to each basic sample; Based on the statement logical relationship information corresponding to each basic sample, the set of related statements and the statement to be analyzed are combined to obtain statement dependency chain information.

6. The training method for the game intent recognition model according to claim 2, characterized in that, The intention decomposition of the statement to be analyzed, based on the target semantics and statement dependency chain information corresponding to each candidate term of each base sample, yields the intention structure information corresponding to each base sample, including: Based on the target semantics and statement dependency chain information corresponding to each candidate term of each base sample, a candidate intent list corresponding to each base sample is generated. For each of the basic samples, perform intent distribution prediction on the candidate intent list to obtain the intent probability distribution for each of the basic samples; Based on the intent probability distribution corresponding to each of the basic samples, the semantic unit decomposition of the statement to be analyzed is performed to obtain the intent structure information corresponding to each of the basic samples.

7. The training method for the game intent recognition model according to any one of claims 1-6, characterized in that, The process of supervising and fine-tuning the pre-trained lightweight language model for different intent tasks based on each input sample and the intent annotation information corresponding to each input sample to obtain the trained game intent recognition model includes: Multiple input samples are sampled according to the task sampling rate corresponding to different intention tasks to obtain multiple target input samples for the current training batch; The multiple target input samples are respectively input into the pre-trained lightweight language model to obtain the intent prediction result corresponding to each target input sample; Based on the intent prediction results and corresponding intent annotation information of each target input sample, loss calculation is performed to obtain the loss values ​​corresponding to different intent tasks. The parameters of the lightweight language model are then updated based on the loss values ​​corresponding to the different intent tasks until the loss converges to obtain the trained game intent recognition model.

8. A training device for a game intent recognition model, characterized in that, include: The acquisition module is configured to acquire the basic sample set; The parsing module is configured to call a pre-trained large language model to perform deep semantic parsing on the data in each basic sample in the basic sample set, and obtain the full intent annotation information corresponding to each basic sample. The splicing module is configured to splice multiple instruction templates corresponding to different pre-generated intent tasks with each of the basic samples to obtain multiple input samples corresponding to each intent task; The decomposition module is configured to decompose the full intent annotation information corresponding to each of the basic samples to obtain the intent annotation information corresponding to each of the input samples; The training module is configured to perform supervised fine-tuning training of a pre-trained lightweight language model for different intent tasks based on each input sample and the intent annotation information corresponding to each input sample, so as to obtain a trained game intent recognition model. The game intent recognition model is used to output the intent recognition results of game dialogue.

9. An electronic device, characterized in that, The device includes: one or more processors; and a storage device configured to store one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement the training method of the game intent recognition model according to any one of claims 1-7.

10. A non-volatile storage medium for storing computer-executable instructions, characterized in that, The computer-executable instructions, when executed by a computer processor, are configured to perform the training method of the game intent recognition model according to any one of claims 1-7.

Images