Multi-task modeling method in game content search association field

Abstract

The invention provides a multi-task modeling method in a game content search association field. The method comprises the following steps: constructing a multi-task modeling model; the multi-task modeling model comprises an input layer, an embedding model, a vector splicing model, an expert network layer, a gating network layer, a gating weighted fusion layer, a multi-task tower layer and an output layer; obtaining a training sample set; each training sample in the training sample set comprises a user current search term query, a user historical search term sequence, a user basic feature set and a user real behavior tag set; and training the multi-task modeling model by adopting the training sample set. According to the method, multi-task modeling can be carried out in the field of game content search association, so that the click preference of the user for the candidate words can be predicted more accurately, and the conversion efficiency of subsequent content consumption is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer information processing, and in particular is a method for multi-task modeling in the field of game content search association. Background Art

[0002] In the field of game content search association, deep neural network models (hereinafter referred to as "models") are widely used to model users' click preferences for recommended candidate search terms (hereinafter referred to as "candidate terms"), thereby determining the order in which candidate terms are recommended to users.

[0003] However, in many cases, users are not satisfied with the search results obtained by clicking on candidate words. Generally, there are several reasons why users are dissatisfied with the results: one is that the ranking of candidate words is not accurate enough. Although the user clicks on the candidate word at the top, the candidate word is not what the user really wants to search for, but there is no better candidate word to choose from. Another reason is that the user does not like the content searched out by the corresponding candidate word. It may be that the content of the platform itself is not rich enough, or the ranking of the content itself is not accurate enough.

[0004] In the prior art, in order to improve the user search experience, modeling is performed based on the click behavior of candidate words, which has the following defects or shortcomings:

[0005] 1. Modeling the target separately has low modeling efficiency and requires high memory and computing power.

[0006] 2. Ignoring that user conversion is a chain, the later the part in the chain, the greater the contribution to the final goal, but the corresponding number of samples is also relatively small. Training the model alone will lead to poor training results due to sparse samples.

[0007] 3. In traditional single-task learning, the model may be easily disturbed by task-specific data, resulting in negative transfer effects. Summary of the invention

[0008] In view of the defects of the prior art, the present invention provides a method for multi-task modeling in the field of game content search association, which can effectively solve the above problems.

[0009] The technical solution adopted by the present invention is as follows:

[0010] The present invention provides a method for multi-task modeling in the field of game content search association, comprising the following steps:

[0011] Step S1, constructing a multi-task modeling model; the multi-task modeling model includes an input layer, an embedding model Embedding, a vector splicing model, an expert network layer, a gated network layer, a gated weighted fusion layer, a multi-task tower layer and an output layer;

[0012] Step S2, obtaining a training sample set; each training sample in the training sample set includes a user's current search query, a user's historical search sequence, a user's basic feature set, and a user's real behavior label set;

[0013] Step S3, using the training sample set to train the multi-task modeling model to obtain the trained multi-task modeling model;

[0014] Step S4, using the trained multi-task modeling model, predicting the probability of various user behaviors based on the user's current search term query in real time.

[0015] Preferably, in step S2, the user history search term sequence includes L user history search terms query_his; the user basic feature set includes a plurality of user basic features base_info corresponding to the user current search term query.

[0016] Preferably, the user basic features base_info include user network, model, application version and forum.

[0017] Preferably, step S3 specifically comprises:

[0018] Step S3.1, inputting the user's current search term query, the user's historical search term sequence and the user's basic feature set into the input layer;

[0019] Step S3.2, through the input layer, the embedding model Embedding and the vector concatenation model, feature fusion is performed on the user's current search term query, the user's historical search term sequence and the user's basic feature set to obtain a feature representation fusion vector hidden_emb;

[0020] Step S3.3, inputting the feature representation fusion vector hidden_emb into the expert network layer; the expert network layer includes n expert networks, each of which performs expert prediction on the feature representation fusion vector hidden_emb and outputs an expert prediction vector; therefore, the output representation vector of the expert network layer is expert_out_en=[out_en_1, out_en_2,..., out_en_n]; wherein out_en_i represents the expert prediction vector of the i-th expert network, i=1,2,…,n;

[0021] Step S3.4, input the feature representation fusion vector hidden_emb to the gated network layer; the gated network layer includes m gated networks, each of which performs gated control on the feature representation fusion vector hidden_emb and outputs a gated control vector; therefore, the output representation vector of the gated network layer is expert_out_gate = [out_gate_1, out_gate_2, ..., out_gate_m]; wherein out_gate_j represents the gated control vector of the jth gated network, j = 1, 2, ..., m;

[0022] Step S3.5, the output representation vector expert_out_en of the expert network layer and the output representation vector expert_out_gate of the gated network layer are both input to the gated weighted fusion layer, and the gated weighted fusion layer executes formula (1):

[0023]

[0024] Specifically, formula (1) is used to perform weighted fusion on each expert prediction vector in the output representation vector expert_out_en of the expert network layer using the output representation vector expert_out_gate output by the gated network layer to achieve gated filtering. The output representation vector multi_out of the gated weighted fusion layer is [g_out_1, g_out_2, ..., g_out_m]; wherein g_out_j represents the gated weighted fusion vector obtained after the gated control vector out_gate_j of the jth gated network performs weighted fusion on the output representation vector expert_out_en of the expert network layer.

[0025] Step S3.6, inputting the output representation vector multi_out of the gated weighted fusion layer into the multi-task tower layer;

[0026] The multi-task tower layer includes m task towers. For the j-th task tower, j=1, 2, ..., m, the corresponding task is executed in the following manner: forward propagation is performed on the j-th gated weighted fusion vector g_out_j, the corresponding task is executed, and the prediction result of the corresponding task is obtained;

[0027] Therefore, the multi-task tower layer outputs a prediction result set, the prediction result set includes m prediction results, and is transmitted to the output layer;

[0028] Step S3.7, the output layer performs loss calculation on each prediction result in the prediction result set and the corresponding user real behavior label in the user real behavior label set, calculates the gradient according to the loss and performs back propagation, thereby updating the model parameters of the multi-task modeling model;

[0029] Step S3.8, using steps S3.1 to S3.7, using each of the training samples in the training sample set, cyclically train the multi-task modeling model until the multi-task modeling model converges, thereby obtaining the trained multi-task modeling model.

[0030] Preferably, in step S3.2, the embedding model Embedding is used to map the input text into a vector representation of length D and randomly initialize the value of the vector representation.

[0031] Preferably, step S3.2 is specifically:

[0032] Step S3.2.1, performing multi-granular semantic representation, average pooling operation and vector concatenation on the user's current search term query to obtain the user's current search term feature vector query_emb;

[0033] Step S3.2.2, using the user's current search term query, weightedly fuses each user's historical search term query_his in the user's historical search term sequence to obtain a user's historical search term feature vector query_his_emb;

[0034] Step S3.2.3, performing vector representation and vector concatenation on each user basic feature base_info in the user basic feature set to obtain a user basic feature vector base_emb;

[0035] Step S3.2.4, using formula (2), concatenate the user's current search term feature vector query_emb, the user's historical search term feature vector query_his_emb, and the user's basic feature vector base_emb to obtain the feature representation fusion vector hidden_emb:

[0036] concat([query_emb,query_his_emb,base_emb])(2)

[0037] Step S3.2 ends the process.

[0038] Preferably, step S3.2.1 is specifically:

[0039] ① Converting the user's current search query into user's current search query of different granularities, wherein the granularity includes single word, double word and word segmentation;

[0040] ② Mapping the user's current search term query at each granularity into a vector representation of the corresponding granularity through the embedding model;

[0041] ③ Perform an average pooling operation on the vector representations of all granularities to reduce the dimension of the vector representation of each granularity, and obtain three reduced-dimensional vector representations corresponding to single-word, double-word, and word segmentation granularities;

[0042] ④ The three vector representations after dimensionality reduction are concatenated using the vector concatenation model to obtain the user's current search term feature vector query_emb with a dimension of 3*D.

[0043] Preferably, step S3.2.2 is specifically:

[0044] ① The user's current search term query is mapped to the user's current search term vector representation u_query through the embedding model Embedding;

[0045] ② The user history search word sequence includes L user history search words query_his; for the k-th user history search word, it is represented as: query_his_k, k=1,2,…,L, and is mapped to the k-th user history search word vector representation u_query_his_k through the embedding model Embedding;

[0046] The k-th user's historical search term vector representation u_query_his_k and the user's current search term vector representation u_query perform the click attention mechanism Dot Product Attention to generate the weight vector att_w_k of the k-th user's historical search term;

[0047] ③ Using formula (3), using the weight vector of each user's historical search term, perform weighted fusion on the sequence represented by the user's historical search term vector to obtain the user's historical search term feature vector query_his_emb;

[0048]

[0049] The process of step S3.2.2 ends.

[0050] Preferably, step S3.2.3 is specifically:

[0051] Each user basic feature base_info in the user basic feature set is mapped to a corresponding vector representation through the embedding model Embedding; and then the mapped vector representations are concatenated to obtain a user basic feature vector base_emb.

[0052] Preferably, the multi-task tower layer includes two task towers, corresponding to tasks: a task of predicting whether a user clicks on a candidate word, and a task of predicting whether a user consumes the content after the search;

[0053] The prediction obtains the probabilities of various user behaviors, specifically: predicting the probability of whether the user clicks on each candidate word, and predicting the probability of the user consuming the content after the search.

[0054] The method for multi-task modeling in the field of game content search association provided by the present invention has the following advantages:

[0055] The present invention provides a method for multi-task modeling in the field of game content search association, which can realize multi-task modeling in the field of game content search association, so as to more accurately predict the user's click preference for candidate words and improve the conversion efficiency of subsequent content consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A flowchart of a method for multi-task modeling in the field of game content search association provided by the present invention;

[0057] Figure 2 This is the expert network calculation flow chart provided by the present invention. DETAILED DESCRIPTION

[0058] In order to make the technical problems, technical solutions and beneficial effects solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0059] In order to solve the related problems existing in the prior art, the present invention provides a method for multi-task modeling in the field of game content search association, which can realize multi-task modeling in the field of game content search association, so as to more accurately predict the user's click preference for candidate words and improve the conversion efficiency of subsequent content consumption.

[0060] See also Figure 1 and Figure 2 The present invention provides a method for multi-task modeling in the field of game content search association, comprising the following steps:

[0061] Step S1, constructing a multi-task modeling model; the multi-task modeling model includes an input layer, an embedding model Embedding, a vector splicing model, an expert network layer, a gated network layer, a gated weighted fusion layer, a multi-task tower layer and an output layer;

[0062] Step S2, obtaining a training sample set; each training sample in the training sample set includes a user's current search query, a user's historical search sequence, a user's basic feature set, and a user's real behavior label set;

[0063] Step S3, using the training sample set to train the multi-task modeling model to obtain the trained multi-task modeling model;

[0064] Step S4, using the trained multi-task modeling model, predicting the probability of various user behaviors based on the user's current search term query in real time.

[0065] The following is a detailed description of steps S2 to S4:

[0066] Step S2, obtaining a training sample set; each training sample in the training sample set includes a user's current search query, a user's historical search sequence, a user's basic feature set, and a user's real behavior label set;

[0067] In this step, the user history search term sequence includes L user history search terms query_his;

[0068] The user basic feature set includes a plurality of user basic features base_info corresponding to the user's current search term query. For example, the user basic feature base_info includes the user's network, model, application version, and forum to which the user belongs.

[0069] Step S3, using the training sample set to train the multi-task modeling model to obtain the trained multi-task modeling model;

[0070] Step S3 is specifically as follows:

[0071] Step S3.1, inputting the user's current search term query, the user's historical search term sequence and the user's basic feature set into the input layer;

[0072] Step S3.2, through the input layer, the embedding model Embedding and the vector concatenation model, feature fusion is performed on the user's current search term query, the user's historical search term sequence and the user's basic feature set to obtain a feature representation fusion vector hidden_emb;

[0073] The embedding model Embedding is used to map the input text into a vector representation with a length of D and randomly initialize the value of the vector representation.

[0074] Step S3.2 is specifically as follows:

[0075] Step S3.2.1, performing multi-granular semantic representation, average pooling operation and vector concatenation on the user's current search term query to obtain the user's current search term feature vector query_emb;

[0076] The specific steps are:

[0077] ① Converting the user's current search query into user's current search query of different granularities, wherein the granularity includes single word, double word and word segmentation;

[0078] ② Mapping the user's current search term query at each granularity into a vector representation of the corresponding granularity through the embedding model;

[0079] ③ Perform an average pooling operation on the vector representations of all granularities to reduce the dimension of the vector representation of each granularity, and obtain three reduced-dimensional vector representations corresponding to single-word, double-word, and word segmentation granularities;

[0080] ④ The three vector representations after dimensionality reduction are concatenated using the vector concatenation model to obtain the user's current search term feature vector query_emb with a dimension of 3*D.

[0081] Step S3.2.2, using the user's current search term query, weightedly fuses each user's historical search term query_his in the user's historical search term sequence to obtain a user's historical search term feature vector query_his_emb;

[0082] The specific steps are:

[0083] ① The user's current search term query is mapped to the user's current search term vector representation u_query through the embedding model Embedding;

[0084] ② The user history search word sequence includes L user history search words query_his; for the k-th user history search word, it is represented as: query_his_k, k=1,2,…,L, and is mapped to the k-th user history search word vector representation u_query_his_k through the embedding model Embedding;

[0085] The k-th user's historical search term vector representation u_query_his_k and the user's current search term vector representation u_query perform the click attention mechanism Dot Product Attention to generate the weight vector att_w_k of the k-th user's historical search term;

[0086] ③ Using formula (3), using the weight vector of each user's historical search term, perform weighted fusion on the sequence represented by the user's historical search term vector to obtain the user's historical search term feature vector query_his_emb;

[0087]

[0088] The process of step S3.2.2 ends.

[0089] Step S3.2.3, performing vector representation and vector concatenation on each user basic feature base_info in the user basic feature set to obtain a user basic feature vector base_emb;

[0090] Step S3.2.3 is specifically:

[0091] Each user basic feature base_info in the user basic feature set is mapped to a corresponding vector representation through the embedding model Embedding; and then the mapped vector representations are concatenated to obtain a user basic feature vector base_emb.

[0092] Step S3.2.4, using formula (2), concatenate the user's current search term feature vector query_emb, the user's historical search term feature vector query_his_emb, and the user's basic feature vector base_emb to obtain the feature representation fusion vector hidden_emb:

[0093] concat([query_emb,query_his_emb,base_emb])(2)

[0094] Step S3.2 ends the process.

[0095] Step S3.3, inputting the feature representation fusion vector hidden_emb into the expert network layer; the expert network layer includes n expert networks, each of which performs expert prediction on the feature representation fusion vector hidden_emb and outputs an expert prediction vector; therefore, the output representation vector of the expert network layer is expert_out_en=[out_en_1, out_en_2,..., out_en_n]; wherein out_en_i represents the expert prediction vector of the i-th expert network, i=1,2,…,n;

[0096] Step S3.4, input the feature representation fusion vector hidden_emb to the gated network layer; the gated network layer includes m gated networks, each of which performs gated control on the feature representation fusion vector hidden_emb and outputs a gated control vector; therefore, the output representation vector of the gated network layer is expert_out_gate = [out_gate_1, out_gate_2, ..., out_gate_m]; wherein out_gate_j represents the gated control vector of the jth gated network, j = 1, 2, ..., m;

[0097] Step S3.5, the output representation vector expert_out_en of the expert network layer and the output representation vector expert_out_gate of the gated network layer are both input to the gated weighted fusion layer, and the gated weighted fusion layer executes formula (1):

[0098]

[0099] Specifically, formula (1) is used to perform weighted fusion on each expert prediction vector in the output representation vector expert_out_en of the expert network layer using the output representation vector expert_out_gate output by the gated network layer to achieve gated filtering. The output representation vector multi_out of the gated weighted fusion layer is [g_out_1, g_out_2, ..., g_out_m]; wherein g_out_j represents the gated weighted fusion vector obtained after the gated control vector out_gate_j of the jth gated network performs weighted fusion on the output representation vector expert_out_en of the expert network layer.

[0100] Step S3.6, inputting the output representation vector multi_out of the gated weighted fusion layer into the multi-task tower layer;

[0101] The multi-task tower layer includes m task towers. For the j-th task tower, j=1, 2, ..., m, the corresponding task is executed in the following manner: forward propagation is performed on the j-th gated weighted fusion vector g_out_j, the corresponding task is executed, and the prediction result of the corresponding task is obtained;

[0102] Therefore, the multi-task tower layer outputs a prediction result set, the prediction result set includes m prediction results, and is transmitted to the output layer;

[0103] As an embodiment, the multi-task tower layer includes two task towers, corresponding to tasks: a task of predicting whether a user clicks on a candidate word, and a task of predicting whether a user consumes the content after the search;

[0104] Step S3.7, the output layer performs loss calculation on each prediction result in the prediction result set and the corresponding user real behavior label in the user real behavior label set, calculates the gradient according to the loss and performs back propagation, thereby updating the model parameters of the multi-task modeling model;

[0105] Step S3.8, using steps S3.1 to S3.7, using each of the training samples in the training sample set, cyclically train the multi-task modeling model until the multi-task modeling model converges, thereby obtaining the trained multi-task modeling model.

[0106] Step S4, using the trained multi-task modeling model, predicts the probabilities of various user behaviors based on the user's current search query in real time, for example, predicting the probability of whether the user clicks on each candidate word, and predicting the probability of the user consuming the content after the search.

[0107] Therefore, through a method for multi-task modeling in the field of game content search association provided by the present invention, for the user's current search word query, considering the user's historical search word sequence and the user's basic feature set, it is possible to obtain multiple probabilities for each candidate word in the candidate word list, including but not limited to the probability of whether the user clicks the candidate word, and the probability of the user consuming the searched content after clicking the candidate word, and then weighting the multiple probabilities obtained to obtain a total probability; for the total probability of each candidate word, adjust the order of each candidate word in the candidate word list, and put the candidate words with a high total probability in front, so as to optimize the order of each candidate word in the candidate word list, improve the user's search experience, and improve the satisfaction of the search results.

[0108] In traditional solutions, single-task modeling is performed, and usually only candidate word clicks are modeled, without taking into account the user's subsequent content consumption behavior, resulting in a low user search experience. In the present invention, candidate word clicks and user's subsequent content consumption behavior are modeled simultaneously. Since user conversion is a chain, the later the part in the chain contributes more to the final goal. Therefore, taking into account the user's subsequent content consumption behavior at the end of the chain can effectively improve the user's search experience and improve the satisfaction of the search results. Furthermore, the present invention adopts a multi-task modeling method to simultaneously model candidate word clicks and user's subsequent content consumption behavior. Multi-task learning can alleviate the negative transfer problem and improve the robustness of the overall model through the joint optimization of multiple tasks.

[0109] Compared with the prior art, the present invention has the following beneficial effects:

[0110] 1) The present invention solves the limitation problem of single-task modeling. The effects of related tasks are mutually enhanced. By learning multiple tasks at the same time, if one task contains information useful for another task, the prediction performance of the latter task can be improved.

[0111] 2) Compared with single-task modeling, the present invention improves the generalization of the model and has a regularization effect, that is, the model needs to perform well not only on one task, but also on other tasks, and tends to learn features that perform well on multiple tasks, making the model more robust;

[0112] 3) The present invention alleviates the low efficiency problem of single-task modeling. Multi-task models can share some structures, reduce memory usage, reduce repeated calculations during reasoning, and improve reasoning speed.

[0113] 4) The present invention can effectively improve the user search experience and enhance the satisfaction of search results through multi-task modeling.

[0114] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be considered as the scope of protection of the present invention.

Claims

1. A method for multi-task modeling in the field of game content search association, characterized in that: The following steps are involved: Step S1, constructing a multi-task modeling model; the multi-task modeling model includes an input layer, an embedding model Embedding, a vector splicing model, an expert network layer, a gated network layer, a gated weighted fusion layer, a multi-task tower layer and an output layer; Step S2, obtaining a training sample set; each training sample in the training sample set includes a user's current search query, a user's historical search sequence, a user's basic feature set, and a user's real behavior label set; Step S3, using the training sample set to train the multi-task modeling model to obtain the trained multi-task modeling model; Step S4, using the trained multi-task modeling model, predicting the probability of various user behaviors based on the user's current search term query in real time.

2. A method for multi-task modeling in the field of game content search association according to claim 1, characterized in that: In step S2, the user history search word sequence includes L user history search words query_his; the user basic feature set includes multiple user basic features base_info corresponding to the user's current search word query.

3. The method for multi-task modeling in the field of game content search association according to claim 1, characterized in that: The user basic features base_info include user network, model, application version and forum.

4. The method for multi-task modeling in the field of game content search association according to claim 1, characterized in that: Step S3 is specifically as follows: Step S3.1, inputting the user's current search term query, the user's historical search term sequence and the user's basic feature set into the input layer; Step S3.2, performing feature fusion on the user's current search term query, the user's historical search term sequence, and the user's basic feature set through the input layer, the embedding model Embedding, and the vector concatenation model to obtain a feature representation fusion vector hidden_emb; Step S3.3, inputting the feature representation fusion vector hidden_emb into the expert network layer; the expert network layer includes n expert networks, each of which performs expert prediction on the feature representation fusion vector hidden_emb and outputs an expert prediction vector; therefore, the output representation vector of the expert network layer is expert_out_en=[out_en_1, out_en_2,..., out_en_n]; wherein out_en_i represents the expert prediction vector of the i-th expert network, i=1,2,…,n; Step S3.4, input the feature representation fusion vector hidden_emb to the gated network layer; the gated network layer includes m gated networks, each of which performs gated control on the feature representation fusion vector hidden_emb and outputs a gated control vector; therefore, the output representation vector of the gated network layer is expert_out_gate = [out_gate_1, out_gate_2, ..., out_gate_m]; wherein out_gate_j represents the gated control vector of the jth gated network, j = 1, 2, ..., m; Step S3.5, the output representation vector expert_out_en of the expert network layer and the output representation vector expert_out_gate of the gated network layer are both input to the gated weighted fusion layer, and the gated weighted fusion layer executes formula (1): Specifically, formula (1) is used to perform weighted fusion on each expert prediction vector in the output representation vector expert_out_en of the expert network layer using the output representation vector expert_out_gate output by the gated network layer to achieve gated filtering. The output representation vector multi_out of the gated weighted fusion layer is [g_out_1, g_out_2, ..., g_out_m]; wherein g_out_j represents the gated weighted fusion vector obtained after the gated control vector out_gate_j of the jth gated network performs weighted fusion on the output representation vector expert_out_en of the expert network layer. Step S3.6, inputting the output representation vector multi_out of the gated weighted fusion layer into the multi-task tower layer; The multi-task tower layer includes m task towers. For the j-th task tower, j=1, 2, ..., m, the corresponding task is executed in the following manner: forward propagation is performed on the j-th gated weighted fusion vector g_out_j, the corresponding task is executed, and the prediction result of the corresponding task is obtained; Therefore, the multi-task tower layer outputs a prediction result set, the prediction result set includes m prediction results, and is transmitted to the output layer; Step S3.7, the output layer performs loss calculation on each prediction result in the prediction result set and the corresponding user real behavior label in the user real behavior label set, calculates the gradient according to the loss and performs back propagation, thereby updating the model parameters of the multi-task modeling model; Step S3.8, using steps S3.1 to S3.7, using each of the training samples in the training sample set, cyclically train the multi-task modeling model until the multi-task modeling model converges, thereby obtaining the trained multi-task modeling model.

5. The method for multi-task modeling in the field of game content search association according to claim 4, characterized in that: In step S3.2, the embedding model Embedding is used to map the input text into a vector representation of length D and randomly initialize the value of the vector representation.

6. The method for multi-task modeling in the field of game content search association according to claim 5, characterized in that: Step S3.2 is specifically as follows: Step S3.2.1, performing multi-granular semantic representation, average pooling operation and vector concatenation on the user's current search term query to obtain the user's current search term feature vector query_emb; Step S3.2.2, using the user's current search term query, weightedly fuses each user's historical search term query_his in the user's historical search term sequence to obtain a user's historical search term feature vector query_his_emb; Step S3.2.3, performing vector representation and vector concatenation on each user basic feature base_info in the user basic feature set to obtain a user basic feature vector base_emb; Step S3.2.4, using formula (2), concatenate the user's current search term feature vector query_emb, the user's historical search term feature vector query_his_emb, and the user's basic feature vector base_emb to obtain the feature representation fusion vector hidden_emb: concat([query_emb,query_his_emb,base_emb])(2) Step S3.2 The process ends.

7. A method for multi-task modeling in the field of game content search association according to claim 6, characterized in that: Step S3.2.1 is specifically as follows: ① Converting the user's current search query into user's current search query of different granularities, wherein the granularity includes single word, double word and word segmentation; ② Mapping the user's current search term query at each granularity into a vector representation of the corresponding granularity through the embedding model; ③ Perform an average pooling operation on the vector representations of all granularities to reduce the dimension of the vector representation of each granularity, and obtain three reduced-dimensional vector representations corresponding to single-word, double-word, and word segmentation granularities; ④ The three vector representations after dimensionality reduction are concatenated using the vector concatenation model to obtain the user's current search term feature vector query_emb with a dimension of 3*D.

8. The method for multi-task modeling in the field of game content search association according to claim 6, characterized in that: Step S3.2.2 is specifically: ① The user's current search term query is mapped to the user's current search term vector representation u_query through the embedding model Embedding; ② The user history search word sequence includes L user history search words query_his; for the k-th user history search word, it is represented as: query_his_k, k=1,2,…,L, and is mapped to the k-th user history search word vector representation u_query_his_k through the embedding model Embedding; The k-th user's historical search term vector representation u_query_his_k and the user's current search term vector representation u_query perform the click attention mechanism Dot Product Attention to generate the weight vector att_w_k of the k-th user's historical search term; ③ Using formula (3), using the weight vector of each user's historical search term, perform weighted fusion on the sequence represented by the user's historical search term vector to obtain the user's historical search term feature vector query_his_emb; The process of step S3.2.2 ends.

9. The method for multi-task modeling in the field of game content search association according to claim 6, characterized in that: Step S3.2.3 is specifically: Each user basic feature base_info in the user basic feature set is mapped to a corresponding vector representation through the embedding model Embedding; and then the mapped vector representations are concatenated to obtain a user basic feature vector base_emb.

10. The method for multi-task modeling in the field of game content search association according to claim 4, characterized in that: The multi-task tower layer includes two task towers, corresponding to tasks: a task of predicting whether a user clicks on a candidate word, and a task of predicting whether a user consumes the content after the search; The prediction obtains the probabilities of various user behaviors, specifically: predicting the probability of whether the user clicks on each candidate word, and predicting the probability of the user consuming the content after the search.

Images