Method for multi-granularity semantic understanding in game content search association field

Abstract

The invention provides a method for multi-granularity semantic understanding in the field of game content search association. The method comprises the following steps: establishing a vector library; the method comprises the following steps: obtaining a single-word granularity average pooling vector q1gram, a double-word granularity average pooling vector q2gram, a word granularity average pooling vector qseg, a historical search word fusion vector qhisseq, a self-attention coding vector qseqtrm and a mixed vector qmix of a search word query, and splicing to obtain a final semantic vector Qmsg of the search word query. The invention provides a method for multi-granularity semantic understanding in the field of game content search association, which can extract multi-granularity semantic expression of search terms in the field of game content search association so as to more accurately predict the current click preference of a user for candidate words.

Description

Technical Field

[0001] The present invention belongs to the field of information processing technology, and in particular relates to a method for multi-granularity semantic understanding in the field of game content search association. Background Art

[0002] In the field of game content search association, deep neural network models are widely used to parse the search terms currently entered by the user, obtain the user's intention, and thus obtain the click probability for different candidate search terms, and then recommend several candidate search terms with the highest click probability to the user based on the click probability.

[0003] However, the existing deep neural network model has the following shortcomings when predicting click probability: it takes the search term currently input by the user as a whole, and is unable to fully understand the semantics of the search term currently input by the user, thereby failing to accurately understand the user's intention, reducing the accuracy of click probability prediction. Summary of the invention

[0004] In view of the defects of the prior art, the present invention provides a method for multi-granularity semantic understanding in the field of game content search association, which can effectively solve the above problems.

[0005] The technical solution adopted by the present invention is as follows:

[0006] The present invention provides a method for multi-granularity semantic understanding in the field of game content search association, comprising the following steps:

[0007] Step S1, establishing a vector library;

[0008] The method for establishing the vector library is as follows: various search terms in the field of game content search are collected, each search term is split into different granularities to form search terms and search words of various granularities, the values ​​of each search term and search word are randomly initialized to obtain corresponding vectors, each vector has a uniform length D, thereby constructing a vector library;

[0009] Step S2, parsing the search term query currently input by the user according to the vector library to obtain a single-word granularity average pooling vector q1_gram;

[0010] Step S3, parsing the search term query currently input by the user according to the vector library to obtain a double-word granularity average pooling vector q2_gram;

[0011] Step S4, parsing the search term query currently input by the user according to the vector library to obtain a word granularity average pooling vector q_seg;

[0012] The specific method is:

[0013] The search term query is segmented, and each segmented word is mapped to a corresponding word vector according to the vector library, thereby generating a word vector list g = [g1, g2, ..., g r ]; g1,g2,…,g r Each word vector is obtained after the search word query is converted;

[0014] For the word vector list g = [g1, g2, ..., g r ] are averagely pooled to obtain the word-granularity average pooled vector q_seg;

[0015] Step S5, comprehensively analyzing the user's search term query and multiple historical search terms of the same user to obtain a historical search term fusion vector q_his_seq;

[0016] Step S6, using the self-attention encoding module, the word vector list g = [g1, g2, ..., g r ] is self-attention encoded to form a word vector matrix formed by all word vectors in the word vector matrix, and a self-attention encoding vector q_seq_trm is obtained; wherein the self-attention encoding module has a self-attention encoding parameter;

[0017] Step S7, adding the single-word granularity average pooling vector q1_gram, the double-word granularity average pooling vector q2_gram, the word granularity average pooling vector q_seg, the historical search word fusion vector q_his_seq and the self-attention encoding vector q_seq_trm to obtain a mixed vector q_mix;

[0018] Step S8, concatenate the single-word granularity average pooling vector q1_gram, the double-word granularity average pooling vector q2_gram, the word granularity average pooling vector q_seg, the historical search word fusion vector q_his_seq, the self-attention encoding vector q_seq_trm and the mixed vector q_mix to obtain the final semantic vector Q_msg of the search word query;

[0019] Step S9, inputting the semantic vector Q_msg into the trained neural network model; the neural network model outputs candidate search terms related to the semantic vector Q_msg and the click probability of each candidate search term.

[0020] Preferably, step S2 specifically comprises:

[0021] Step S2.1, convert the search term query into single words of unigram, and map each single word into a corresponding single word vector according to the vector library, thereby generating a single word vector list q = [q1, q2, ..., q n];q1,q2,…,q n Each word vector is obtained after the search term query is converted;

[0022] Step S2.2, for the word vector list q = [q1, q2, ..., q n The single-word vectors in ] are average pooled to obtain the single-word granularity average pooled vector q1_gram.

[0023] Preferably, step S3 specifically comprises:

[0024] Step S3.1, convert the search term query into a 2-gram big word, and map each big word into a corresponding big word vector according to the vector library, thereby generating a big word vector list f = [f1, f2, ..., f m ]; f1,f2,…,f m Each bigram vector obtained after the conversion of the search term query;

[0025] Step S3.2, for the double word vector list f = [f1, f2, ..., f m The double-word vectors in ] are average pooled to obtain the double-word granularity average pooled vector q2_gram.

[0026] Preferably, step S5 is specifically:

[0027] Step S5.1, mapping the search word query as a whole into a search word vector Q' according to the vector library;

[0028] Obtain a set of historical search terms of the same user; map each historical search term in the set of historical search terms into a corresponding historical search term vector according to the vector library, thereby forming a historical search term vector matrix k'=[k1', k2', ..., k t ']; k1', k2', ..., k t 'Search word vectors for each history;

[0029] Step S5.2, make the search word vector Q' respectively with the historical search word vector matrix k'=[k1',k2',...,k t '] perform vector dot product on each historical search word vector to obtain t scores;

[0030] Step S5.3, using the normalized exponential function softmax, normalize the t scores so that the sum of the normalized t scores is 1, and the normalized t scores form the attention weight vector W';

[0031] Step S5.4, use the attention weight vector W' to analyze the historical search word vector matrix k'=[k1',k2',...,kt '] perform weighted fusion to obtain the historical search term fusion vector q_his_seq.

[0032] Preferably, step S6 specifically comprises:

[0033] Step S6.1, the word vector list g obtained in step S4 is g = [g1, g2, ..., g r The word vector matrix formed by all word vectors in ] is represented as vector matrix S';

[0034] Step S6.2, the self-attention encoding parameters of the self-attention encoding module include the query weight matrix W q , key weight matrix W k Sum value weight matrix W v ;

[0035] Let the vector matrix S' and the query weight matrix W q Multiply them together to perform the first linear projection on the vector matrix S' and obtain the query vector matrix Q;

[0036] Let the vector matrix S' and the key weight matrix W k Multiply them together to perform a second linear projection on the vector matrix S' to obtain the key vector matrix K;

[0037] Let the vector matrix S' and the value weight matrix W v Multiply them together to perform the third linear projection on the vector matrix S' and obtain the value vector matrix V;

[0038] Step S6.3, transpose the key vector matrix K to obtain the transposed key vector matrix K T ;

[0039] Step S6.4, use the following formula to make the transposed key vector matrix K T Perform matrix multiplication with the query vector matrix Q and normalize it with the normalized exponential function softmax to obtain the attention weight matrix A':

[0040] A'=softmax(K T Q)

[0041] Step S6.5, multiply the value vector matrix V and the attention weight matrix A' using the following formula to obtain the self-attention output vector matrix O:

[0042] O=VA'

[0043] Step S6.6, add the self-attention output vector matrix O and the vector matrix S' to obtain the updated vector matrix S', return to step S6.2, and repeat this cycle four times, and finally output the search word encoding matrix X;

[0044] Step S6.7, perform average pooling on the search word encoding matrix X to obtain the self-attention encoding vector q_seq_trm.

[0045] The method for multi-granular semantic understanding in the field of game content search association provided by the present invention has the following advantages:

[0046] The present invention provides a method for multi-granular semantic understanding in the field of game content search association, which can extract multi-granular semantic expressions of search terms query in the field of game content search association, thereby more accurately predicting the user's current click preference for candidate terms. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A schematic diagram of the principles of a method for multi-granularity semantic understanding in the field of game content search association provided by the present invention. DETAILED DESCRIPTION

[0048] 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.

[0049] The present invention provides a method for multi-granular semantic understanding in the field of game content search association, which can extract multi-granular semantic expressions of search terms query in the field of game content search association, thereby more accurately predicting the user's current click preference for candidate terms.

[0050] Specifically, the present invention provides a method for multi-granular semantic understanding in the field of game content search association, which has the following characteristics:

[0051] 1) When identifying the user's search intention, the present invention can mine the user's current search term query in detail, accurately and comprehensively, and adopt a multi-granularity semantic query matrix representation, including unigram single-word granularity, 2-gram double-word granularity, word segmentation granularity and transformer sentence granularity, so as to express the search term query in a multi-granularity manner, mine the semantic information of the user's search term query from different levels, and improve the semantic understanding of the user's current search term query.

[0052] 2) When identifying the user's search intention, the present invention comprehensively considers the correlation between the historical search word sequence and the current search word query, screens the personalized signals based on attention, and differentially introduces the user's personalized behavior. By focusing on screening the personalized signals related to the current search word query, the problem of damaging the relevance when introducing personalization can be avoided, thereby improving the comprehensiveness of the semantic understanding of the user's current search word query.

[0053] See also Figure 1 The present invention provides a method for multi-granularity semantic understanding in the field of game content search association, comprising the following steps:

[0054] Step S1, establishing a vector library;

[0055] The method for establishing the vector library is as follows: various search terms in the field of game content search are collected, each search term is split into different granularities to form search terms and search words of various granularities, the values ​​of each search term and search word are randomly initialized to obtain corresponding vectors, each vector has a uniform length D, thereby constructing a vector library;

[0056] Step S2, parsing the search term query currently input by the user according to the vector library to obtain a single-word granularity average pooling vector q1_gram;

[0057] Step S2 is specifically as follows:

[0058] Step S2.1, convert the search term query into single words of unigram, and map each single word into a corresponding single word vector according to the vector library, thereby generating a single word vector list q = [q1, q2, ..., q n ];q1,q2,…,q n Each word vector is obtained after the search term query is converted;

[0059] Step S2.2, for the word vector list q = [q1, q2, ..., q n The single-word vectors in ] are average pooled to obtain the single-word granularity average pooled vector q1_gram.

[0060] For example, if the search term query is "game strategy", it can be converted into four words, namely: "game", "game", "strategy", and "strategy", thereby forming a single-word vector list q consisting of 4 single-word vectors.

[0061] Step S3, parsing the search term query currently input by the user according to the vector library to obtain a double-word granularity average pooling vector q2_gram;

[0062] Step S3 is specifically as follows:

[0063] Step S3.1, convert the search term query into a 2-gram big word, and map each big word into a corresponding big word vector according to the vector library, thereby generating a big word vector list f = [f1, f2, ..., f m ]; f1,f2,…,f m Each bigram vector obtained after the conversion of the search term query;

[0064] Step S3.2, for the double word vector list f = [f1, f2, ..., f m The double-word vectors in ] are average pooled to obtain the double-word granularity average pooled vector q2_gram.

[0065] For example, if the search term query is "game strategy", it can be converted into three big words, namely: "game", "game strategy", and "strategy", thereby forming a big word vector list f consisting of three big word vectors.

[0066] Step S4, parsing the search term query currently input by the user according to the vector library to obtain a word granularity average pooling vector q_seg;

[0067] The specific method is:

[0068] The search term query is segmented using a word segmenter, and each segmented word is mapped to a corresponding word vector according to the vector library, thereby generating a word vector list g = [g1, g2, ..., g r ]; g1,g2,…,g r Each word vector is obtained after the search word query is converted;

[0069] For the word vector list g = [g1, g2, ..., g r ] are averagely pooled to obtain the word-granularity average pooled vector q_seg;

[0070] For example, if the search query is "game strategy", it can be segmented into two words "game" and "strategy", thus forming a word vector list g consisting of two word segmentation vectors.

[0071] Step S5, comprehensively analyzing the user's search term query and multiple historical search terms of the same user to obtain a historical search term fusion vector q_his_seq;

[0072] Step S5 is specifically as follows:

[0073] Step S5.1, mapping the search word query as a whole into a search word vector Q' according to the vector library;

[0074] Obtain a set of historical search terms of the same user; map each historical search term in the set of historical search terms into a corresponding historical search term vector according to the vector library, thereby forming a historical search term vector matrix k'=[k1', k2', ..., k t ']; k1', k2', ..., k t 'Search word vectors for each history;

[0075] Step S5.2, make the search word vector Q' respectively with the historical search word vector matrix k'=[k1',k2',...,k t '] perform vector dot product on each historical search word vector to obtain t scores; the vector dot product refers to the click attention mechanism Dot Product Attention processing on two vector matrices.

[0076] Step S5.3, use the normalized exponential function softmax to normalize the t scores so that the sum of the normalized t scores is 1. The normalized t scores form the attention weight vector W', and the formula is: W'=softmax(Q'·k').

[0077] Step S5.4, use the attention weight vector W' to analyze the historical search word vector matrix k'=[k1',k2',...,k t '] perform weighted fusion to obtain the historical search term fusion vector q_his_seq.

[0078] Step S6, using the self-attention encoding module of the Transformer model, the word vector list g = [g1, g2, ..., g r ] is self-attention encoded to form a word vector matrix formed by all word vectors in the word vector matrix, and a self-attention encoding vector q_seq_trm is obtained; wherein the self-attention encoding module has a self-attention encoding parameter;

[0079] Step S6 is specifically as follows:

[0080] Step S6.1, the word vector list g obtained in step S4 is g = [g1, g2, ..., g r The word vector matrix formed by all word vectors in ] is represented as vector matrix S';

[0081] Step S6.2, the self-attention encoding parameters of the self-attention encoding module include the query weight matrix W q , key weight matrix W k Sum value weight matrix W v ;

[0082] Let the vector matrix S' and the query weight matrix W qMultiply them together to perform the first linear projection on the vector matrix S' and obtain the query vector matrix Q;

[0083] Let the vector matrix S' and the key weight matrix W k Multiply them together to perform a second linear projection on the vector matrix S' to obtain the key vector matrix K;

[0084] Let the vector matrix S' and the value weight matrix W v Multiply them together to perform the third linear projection on the vector matrix S' and obtain the value vector matrix V;

[0085] Step S6.3, transpose the key vector matrix K to obtain the transposed key vector matrix K T ;

[0086] Step S6.4, use the following formula to make the transposed key vector matrix K T Perform matrix multiplication with the query vector matrix Q and normalize it with the normalized exponential function softmax to obtain the attention weight matrix A':

[0087] A'=softmax(K T Q)

[0088] Step S6.5, use the following formula to multiply the value vector matrix V and the attention weight matrix A' to obtain the self-attention output vector matrix O, the formula is: O = VA'.

[0089] Step S6.6, add the self-attention output vector matrix O and the vector matrix S' to obtain the updated vector matrix S', return to step S6.2, and repeat this cycle four times, and finally output the search word encoding matrix X;

[0090] Step S6.7, perform average pooling on the search word encoding matrix X to obtain the self-attention encoding vector q_seq_trm.

[0091] Step S7, adding the single-word granularity average pooling vector q1_gram, the double-word granularity average pooling vector q2_gram, the word granularity average pooling vector q_seg, the historical search word fusion vector q_his_seq and the self-attention encoding vector q_seq_trm to obtain a mixed vector q_mix;

[0092] Step S8, concatenate the single-word granularity average pooling vector q1_gram, the double-word granularity average pooling vector q2_gram, the word granularity average pooling vector q_seg, the historical search word fusion vector q_his_seq, the self-attention encoding vector q_seq_trm and the mixed vector q_mix to obtain the final semantic vector Q_msg of the search word query, the formula is:

[0093] Q_msg=concat([q1_gram,q2_gram,q_seg,q_his_seq,q_seq_trm,q_mix])

[0094] The "concat" function represents the concatenation of vectors. The concatenated vector Q_msg contains semantic representations of user behaviors at various granularities, and the vector length is 6D. Using this vector can enable the neural network model to better learn the user's intentions and improve the model's prediction accuracy.

[0095] Step S9, inputting the semantic vector Q_msg into the trained neural network model; the neural network model outputs candidate search terms related to the semantic vector Q_msg and the click probability of each candidate search term.

[0096] The aforementioned steps S1 to S9 describe the process of understanding and predicting the actual search term query for the trained self-attention encoding module and neural network model.

[0097] The self-attention encoding module and the neural network model initially need to use search word samples for model training. As a method, the model training is:

[0098] Step C1, establishing a vector library;

[0099] The same as step S1, which will not be repeated here;

[0100] Step C2, for a certain search word sample, adopt the method of step S2 to step S5 to obtain the single-word granularity average pooling vector q1_gram_sample, the double-word granularity average pooling vector q2_gram_sample, the word granularity average pooling vector q_seg_sample and the historical search word fusion vector q_his_seq_sample respectively;

[0101] Step C3, construct a self-attention encoding module so that the query weight matrix W q , key weight matrix W k Sum value weight matrix W v Randomization assignment;

[0102] Step C4, using the self-attention encoding module to obtain the self-attention encoding vector q_seq_trm_sample;

[0103] Step C5, obtaining a mixed vector q_mix_sample;

[0104] Step C6, perform vector concatenation to obtain a semantic vector Q_msg_sample;

[0105] Step C7, inputting the semantic vector Q_msg_sample into the trained neural network model, and the neural network model outputs the click probability of the candidate search term, thereby obtaining the loss value according to the click probability;

[0106] According to the loss value, reversely adjust the query weight matrix W of step C3 q , key weight matrix W k Sum value weight matrix W v The values ​​of , and the parameters of the neural network model;

[0107] Step C8, based on the adjusted query weight matrix W q , key weight matrix W k , value weight matrix W v And the neural network model, use the next search word sample, return to step C2 for training, and continue to cycle until the loss value is less than the set value, and output the final query weight matrix W q , key weight matrix W k , value weight matrix W v and neural network models.

[0108] The present invention provides a method for multi-granular semantic understanding in the field of game content search association, which has the following advantages:

[0109] (1) The user's search term query, candidate terms and historical search terms usually exist in the form of text and cannot be directly used for the training and prediction of the neural network model. The present invention converts the search term query, candidate terms and historical search terms into semantic vectors in numerical form, which can be used for the training and prediction of the neural network model.

[0110] (2) The present invention performs semantic mining on the user's search query at different granularities, taking into account the semantic importance of finer-grained words and characters within the search query, thereby improving the semantic understanding of the search query.

[0111] (3) The present invention takes into account the correlation between the user's historical search term sequence and the currently input search term query, and also takes into account the differences between the correlations between different historical search terms and the currently input search term query, thereby realizing the role of differentiated historical behavior in intent prediction and improving the accuracy of understanding the user's current search intent.

[0112] 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-granular semantic understanding in the field of game content search association, characterized in that: The following steps are involved: Step S1, establishing a vector library; The method for establishing the vector library is as follows: various search terms in the field of game content search are collected, each search term is split into different granularities to form search terms and search words of various granularities, the values ​​of each search term and search word are randomly initialized to obtain corresponding vectors, each vector has a uniform length D, thereby constructing a vector library; Step S2, parsing the search term query currently input by the user according to the vector library to obtain a single-word granularity average pooling vector q1_gram; Step S3, parsing the search term query currently input by the user according to the vector library to obtain a double-word granularity average pooling vector q2_gram; Step S4, parsing the search term query currently input by the user according to the vector library to obtain a word granularity average pooling vector q_seg; The specific method is: The search term query is segmented, and each segmented word is mapped to a corresponding word vector according to the vector library, thereby generating a word vector list g = [g1, g2, ..., g r ]; g1,g2,…,g r Each word vector is obtained after the search word query is converted; For the word vector list g = [g1, g2, ..., g r ] are averagely pooled to obtain the word-granularity average pooled vector q_seg; Step S5, comprehensively analyzing the user's search term query and multiple historical search terms of the same user to obtain a historical search term fusion vector q_his_seq; Step S6, using the self-attention encoding module, the word vector list g = [g1, g2, ..., g r ] is self-attention encoded to form a word vector matrix formed by all word vectors in the word vector matrix, and a self-attention encoding vector q_seq_trm is obtained; wherein the self-attention encoding module has a self-attention encoding parameter; Step S7, adding the single-word granularity average pooling vector q1_gram, the double-word granularity average pooling vector q2_gram, the word granularity average pooling vector q_seg, the historical search word fusion vector q_his_seq and the self-attention encoding vector q_seq_trm to obtain a mixed vector q_mix; Step S8, concatenate the single-word granularity average pooling vector q1_gram, the double-word granularity average pooling vector q2_gram, the word granularity average pooling vector q_seg, the historical search word fusion vector q_his_seq, the self-attention encoding vector q_seq_trm and the mixed vector q_mix to obtain the final semantic vector Q_msg of the search word query; Step S9, inputting the semantic vector Q_msg into the trained neural network model; the neural network model outputs candidate search terms related to the semantic vector Q_msg and the click probability of each candidate search term.

2. A method for multi-granular semantic understanding in the field of game content search association according to claim 1, characterized in that: Step S2 is specifically as follows: Step S2.1, convert the search term query into single words of unigram, and map each single word into a corresponding single word vector according to the vector library, thereby generating a single word vector list q = [q1, q2, ..., q n ];q1,q2,…,q n Each word vector is obtained after the search term query is converted; Step S2.2, for the word vector list q = [q1, q2, ..., q n The single-word vectors in ] are average pooled to obtain the single-word granularity average pooled vector q1_gram.

3. A method for multi-granular semantic understanding in the field of game content search association according to claim 1, characterized in that: Step S3 is specifically as follows: Step S3.1, convert the search term query into a 2-gram big word, and map each big word into a corresponding big word vector according to the vector library, thereby generating a big word vector list f = [f1, f2, ..., f m ]; f1,f2,…,f m Each bigram vector obtained after the conversion of the search term query; Step S3.2, for the double word vector list f = [f1, f2, ..., f m The double-word vectors in ] are average pooled to obtain the double-word granularity average pooled vector q2_gram.

4. The method for multi-granular semantic understanding in the field of game content search association according to claim 1, characterized in that: Step S5 is specifically as follows: Step S5.1, mapping the search word query as a whole into a search word vector Q' according to the vector library; Obtain a set of historical search terms of the same user; map each historical search term in the set of historical search terms into a corresponding historical search term vector according to the vector library, thereby forming a historical search term vector matrix k'=[k1', k2', ..., k t ']; k1', k2', ..., k t 'Search word vectors for each history; Step S5.2, make the search word vector Q' respectively with the historical search word vector matrix k'=[k1',k2',...,k t '] perform vector dot product on each historical search word vector to obtain t scores; Step S5.3, using the normalized exponential function softmax, normalize the t scores so that the sum of the normalized t scores is 1, and the normalized t scores form the attention weight vector W'; Step S5.4, use the attention weight vector W' to analyze the historical search word vector matrix k'=[k1',k2',...,k t '] perform weighted fusion to obtain the historical search term fusion vector q_his_seq.

5. The method for multi-granular semantic understanding in the field of game content search association according to claim 1, characterized in that: Step S6 is specifically as follows: Step S6.1, the word vector list g obtained in step S4 is g = [g1, g2, ..., g r The word vector matrix formed by all word vectors in ] is represented as vector matrix S'; Step S6.2, the self-attention encoding parameters of the self-attention encoding module include the query weight matrix W q , key weight matrix W k Sum value weight matrix W v ; Let the vector matrix S' and the query weight matrix W q Multiply them together to perform the first linear projection on the vector matrix S' and obtain the query vector matrix Q; Let the vector matrix S' and the key weight matrix W k Multiply them together to perform a second linear projection on the vector matrix S' to obtain the key vector matrix K; Let the vector matrix S' and the value weight matrix W v Multiply them together to perform the third linear projection on the vector matrix S' and obtain the value vector matrix V; Step S6.3, transpose the key vector matrix K to obtain the transposed key vector matrix K T ; Step S6.4, use the following formula to make the transposed key vector matrix K T Perform matrix multiplication with the query vector matrix Q and normalize it with the normalized exponential function softmax to obtain the attention weight matrix A': A’=softmax(K T Q) Step S6.5, multiply the value vector matrix V and the attention weight matrix A' using the following formula to obtain the self-attention output vector matrix O: O=VA' Step S6.6, add the self-attention output vector matrix O and the vector matrix S' to obtain the updated vector matrix S', return to step S6.2, and repeat this cycle four times, and finally output the search word encoding matrix X; Step S6.7, perform average pooling on the search word encoding matrix X to obtain the self-attention encoding vector q_seq_trm.

Images