CN111274359B - Query recommendation method and system based on improved VHRED and reinforcement learning - Google Patents

Query recommendation method and system based on improved VHRED and reinforcement learning Download PDF

Info

Publication number
CN111274359B
CN111274359B CN202010067232.4A CN202010067232A CN111274359B CN 111274359 B CN111274359 B CN 111274359B CN 202010067232 A CN202010067232 A CN 202010067232A CN 111274359 B CN111274359 B CN 111274359B
Authority
CN
China
Prior art keywords
query
user
network
target
session
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202010067232.4A
Other languages
Chinese (zh)
Other versions
CN111274359A (en
Inventor
陈羽中
胡潇炜
郭昆
陈泽林
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Fuzhou University
Original Assignee
Fuzhou University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Fuzhou University filed Critical Fuzhou University
Priority to CN202010067232.4A priority Critical patent/CN111274359B/en
Publication of CN111274359A publication Critical patent/CN111274359A/en
Application granted granted Critical
Publication of CN111274359B publication Critical patent/CN111274359B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/3331—Query processing
    • G06F16/3332—Query translation
    • G06F16/3338—Query expansion
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention relates to a query recommendation method and a query recommendation system based on improved VHRED and reinforcement learning, wherein the method comprises the following steps: step A: collecting user inquiry log records of a search engine, preprocessing the user inquiry log record data, and constructing a user inquiry log training setTS(ii) a And B, step B: training set using user query logsTSTraining a query recommendation deep learning network model based on VHRED with time characteristics and reinforcement learning; and C: and the query recommendation system receives the query sentence input by the user, inputs the query sentence into the trained query recommendation deep learning network model and outputs the matched query recommendation. The method and the system are beneficial to generating the query recommendation meeting the requirements of the user.

Description

Query recommendation method and system based on improved VHRED and reinforcement learning
Technical Field
The invention relates to the field of natural language processing, in particular to a query recommendation method and system based on improved VHRED and reinforcement learning.
Background
Query suggestions provide suggested queries for the session entered by the user. The query suggestions can enable the search engine to better understand the query intent of the user, and thus, the query of the user can be better optimized. Therefore, this task has received considerable attention in the last decade.
Cao et al propose a context-aware query suggestion framework-considering the entire sequence of queries in a session, rather than just the last query. They constructed a concept sequence suffix tree using query clusters for efficient and effective context-aware query suggestion. The query sequence may also be modeled with a mixed variable memory markov model. Context-aware query suggestions consider more user actions in the session, thereby better modeling information needs. Thus, the idea is also valid in terms of query classification and ranking. Ozertem et al developed a ranking framework that learned to suggest queries directly from search behavior in user search logs. It uses large-scale search logs, avoiding the requirement of manual labeling. Supervised advisory systems are generally more accurate and flexible than unsupervised advisory systems. Their suggested results may also improve diversified and personalized searches. Sordoni et al developed a hierarchical codec model (HRED) for context-aware query suggestion. The encoder first encodes query terms into query embedding using a two-stage Recurrent Neural Network (RNN), and then encodes query sequences into session embedding. It then decodes the session embedded in the target proposal. HRED avoids sparsity using a smooth distribution representation, better utilizing the large scale training data available in the search logs. One recent study has upgraded the sequence-to-sequence model of HRED, modeling different query importance and repeated terms in the session using attention and coping mechanisms.
In the former model, only the HRED model is used as a generation model, and the effect of the model still has a promotion space. Moreover, most models ignore the time characteristics of the query, and the time characteristics have great influence on the generation effect of the models. If the query is generated by only using the generator model, the generated query can not be guaranteed to be prepared to be close to the query generated by the user, so that the generated query has obvious machine generation traces and can not well express the query intention of the user.
Disclosure of Invention
The invention aims to provide a query recommendation method and system based on improved VHRED and reinforcement learning, which are beneficial to generating query recommendations meeting the needs of users.
In order to achieve the purpose, the invention adopts the technical scheme that: a query recommendation method based on improved VHRED and reinforcement learning comprises the following steps:
step A: collecting user query log records of a search engine, preprocessing the user query log record data, and constructing a user query log Training Set (TS);
and B: training a query recommendation deep learning network model based on VHRED with time characteristics and reinforcement learning by using a user query log Training Set (TS);
step C: and the query recommendation system receives the query sentence input by the user, inputs the query sentence into the trained query recommendation deep learning network model and outputs the matched query recommendation.
Further, the step a specifically includes the following steps:
step A1: collecting user query log records of a search engine to obtain an original query log set; wherein each query log of the search engine is represented by a triplet (u, q, t), u representing a user, q representing a query, and t representing a query time;
step A2: dividing an original query log set according to users, and sequencing according to query time to obtain query log subsets of different users;
step A3: setting a time interval T according to the following rule: query logs with query time intervals larger than T belong to different sessions, queries in different sessions are not related to each other, the last query in the same session is a target query containing a query intention of a user, a query log subset of each user is further divided into a plurality of sessions to obtain a session set of each user, and the session sets of all the users form a user query log training set TS;
one session of one user u in TS is represented as
Figure BDA0002376338810000021
Wherein q isiIndicating the ith query, t, in the sessioniDenotes qiCorresponding to the query time, the session contains ku+1 queries, last query
Figure BDA0002376338810000022
Is true of the userReal target query;
to q isiAfter word segmentation and stop word removal, q is addediIs further shown as
Figure BDA0002376338810000023
Denotes qiThe j-th word in (1, 2., L (q) ·i)，L(qi) Denotes qiThe number of words of; q. q.siCorresponding query time tiIs denoted by ti＝(xi,yi,zi,di)，xiRepresents hour, yiRepresents minute, ziDenotes second, diIndicating the day of the week.
Further, the query recommendation deep learning network model comprises a generator network based on a variable hierarchical encoder-decoder recurrent neural network with time characteristics VHRED and a discriminator network based on a hierarchical self-encoder, wherein the hierarchical self-encoder encodes words, sentences and paragraphs by multiple layers of GRUs respectively so as to capture semantic structure information of different levels; the step B specifically comprises the following steps:
step B1: inputting query text and query time pairs in a user query log training set TS into a generator network by taking a user session as a unit, and outputting a target query predicted by the generator network;
step B2: calculating the gradient of each parameter in the generator network by using a back propagation method according to the target loss function loss, and updating the parameters by using a random gradient descent method;
step B3: inputting the target query predicted in the step B1 and the real target query of the user in the user session into a discriminator network, outputting category probability, and judging whether the input target query is the target query predicted by the generator network or the real target query of the user according to the category probability;
step B4: taking the class probability output by the discriminator network in the step B3 as the reward of the generator network, and performing reinforcement learning training by using a strategy gradient method to maximize the return expectation;
step B5: and when the iterative change of the loss value generated by inquiring the recommended deep learning network model is smaller than a set threshold value or reaches the maximum iteration times, terminating the training of the inquiring recommended deep learning network model.
Further, the step B1, with the user session as a unit, inputs the query text and the query time pair in the training set TS of the user query log into the generator network, and outputs the target query predicted by the generator network, specifically includes the following steps:
step B11: taking the user session as a unit, and carrying out the text and query time pair (q) on each query except the target query in the user sessioni,ti) Coding to obtain a characterization vector
Figure BDA0002376338810000031
If the user session is
Figure BDA0002376338810000032
A query text to query time pair (q) in the user sessioni,ti) Is characterized by a token vector
Figure BDA0002376338810000033
Is shown as
Figure BDA0002376338810000034
i＝1,2,...,ku，
Figure BDA0002376338810000035
To characterize a vector
Figure BDA0002376338810000036
And
Figure BDA0002376338810000037
the connection of (a) to (b),
Figure BDA0002376338810000038
is qiThe characterization vector of (a) is determined,
Figure BDA0002376338810000039
is tiThe characterization vector of (2);
wherein q isiThe coding formula of (a) is as follows:
Figure BDA00023763388100000310
wherein GRU represents a gated recurrent neural network,
Figure BDA00023763388100000311
j＝1,2,...,L(qi) Is qiThe j-th word in
Figure BDA00023763388100000312
Is used to represent the word vector of (a),
Figure BDA00023763388100000313
by using a pre-trained word vector matrix E ∈ Rd×|D|Where D represents the dimension of the word vector, | D | is the number of words in the lexicon D,
Figure BDA00023763388100000314
tithe coding formula of (a) is as follows:
Figure BDA0002376338810000041
wherein,
Figure BDA0002376338810000042
connection characterization vector
Figure BDA0002376338810000043
And
Figure BDA0002376338810000044
to obtain
Figure BDA0002376338810000045
Step B12: with the user session as a unit, forming a sequence of the characterization vectors of each query text and query time pair except for the target query in the user session, inputting the sequence into an encoder module of a generator network based on a GRU network for encoding to obtain a user session SuIs characterized by a token vector
Figure BDA0002376338810000046
All query information except the target query in the user session is contained;
wherein S isuThe coding formula of (a) is as follows:
Figure BDA0002376338810000047
step B13: obtained according to step B12
Figure BDA0002376338810000048
Computing user sessions SuLatent variable z ofu；
Firstly, the first step is to
Figure BDA0002376338810000049
The mean value mu is obtained by the feedforward neural network module of the input generator networkuThe formula is as follows:
Figure BDA00023763388100000410
wherein
Figure BDA00023763388100000417
dzIs a hidden variable zuTanh is the hyperbolic tangent function, fFNNIs a feedforward neural network;
the mean value muuInputting a softplus function, and calculating to obtain covariance sigmauThe formula is as follows:
∑u＝softplus(f(μu))
wherein,
Figure BDA00023763388100000411
softplus is an activation function, softplus (x) log (1+ e)x)；
Then, according to the mean value muuSum covariance ∑uObtaining the latent variable z by random sampling calculationuThe formula is as follows:
Figure BDA00023763388100000412
wherein samples is a random number vector,
Figure BDA00023763388100000413
extracting d from a standard normal distributionzA plurality of random numbers, constituting a random number vector samples,
Figure BDA00023763388100000414
is a vector sigmauHadamard product with samples to obtain hidden variables of user session
Figure BDA00023763388100000415
Step B14: subjecting the product obtained in step B12
Figure BDA00023763388100000416
And z obtained in step B13uThe GRU network-based decoder module which is input into the generator network decodes and outputs the target query predicted by the generator network;
first, a token vector of a user session is calculated
Figure BDA0002376338810000051
Initial hidden state of
Figure BDA0002376338810000052
The formula is as follows:
Figure BDA0002376338810000053
wherein, W1Is a weight parameter that is a function of,
Figure BDA0002376338810000054
dhis h0Dimension of (b)0Is a bias term;
firstly, h obtained by the above formula0Decoding through GRU network, decoding the decoded hidden state vector through GRU network, repeating KtargetSecondary decoding to generate a block containing KtargetWord-by-word target query qsWherein the decoding formula is as follows:
Figure BDA0002376338810000055
each decoding produces the next word
Figure BDA0002376338810000056
The probability of (c) is:
Figure BDA0002376338810000057
Figure BDA0002376338810000058
wherein f is the full junction layer, W2、W3、bprobParameters of the fully connected layer, bprobIn order to be a term of the offset,
Figure BDA0002376338810000059
W3∈Rd×d，bprob∈Rd；
candidate target query q'sScore of(s) (score of q's) For the probability product of decoding, the formula is as follows:
Figure BDA00023763388100000510
wherein
Figure BDA00023763388100000511
To constitute target query q'sThe word sequence of (a);
taking logarithm of the above formula, the formula is as follows:
Figure BDA00023763388100000512
selecting
Figure BDA00023763388100000513
Highest q'sTarget query q as predicted by a generator networks：
Figure BDA00023763388100000514
Further, in the step B2, the target loss function loss is defined as follows:
loss＝loss1+loss2
among them, loss1The loss value loss is obtained using KL divergence measure as the difference between the distribution of hidden variables and the unit Gaussian distribution1The calculation is as follows:
Figure BDA0002376338810000061
wherein, muu,ΣuThe mean and covariance obtained in step B13;
loss2target query q predicted for a generator networksAnd a session SuTarget query representing real query intention of user
Figure BDA0002376338810000062
By cross entropy loss functionThe loss value obtained by number calculation is calculated as follows:
Figure BDA0002376338810000063
wherein Cross EntropyLoss is a cross entropy loss function;
and updating the learning rate through a gradient optimization algorithm AdaGrad, and updating model parameters through back propagation iteration so as to minimize a loss function to train the model.
Further, the step B3, inputting the target query predicted in the step B1 and the real target query of the user in the user session into the discriminator network, outputting the category probability, and determining whether the input target query is the target query predicted by the generator network or the real target query of the user according to the category probability specifically includes the following steps:
step B31: the user session S obtained by B11uRemoving user's true target query
Figure BDA0002376338810000064
Outer text of each query qi,i＝1,2,...,kuInputting the data into a discriminator network for coding to obtain qiIs characterized by a token vector
Figure BDA0002376338810000065
i＝1,2,...,ku；
Step B32: q obtained in the step B32iIs characterized by a token vector
Figure BDA0002376338810000066
Constituting a sequence
Figure BDA0002376338810000067
The GRU network module in the input discriminator network is coded to obtain
Figure BDA0002376338810000068
The coding formula is as follows:
Figure BDA0002376338810000069
step B33: respectively conversing users SuUser's true target query in
Figure BDA00023763388100000610
Target query q predicted from step B14sThe GRU network module of the input discriminator network is coded to obtain
Figure BDA00023763388100000611
And q issIs characterized by a token vector
Figure BDA00023763388100000612
And
Figure BDA00023763388100000613
the coding formula is as follows:
Figure BDA00023763388100000614
Figure BDA00023763388100000615
wherein,
Figure BDA00023763388100000616
for querying
Figure BDA00023763388100000617
Middle j (th) word
Figure BDA00023763388100000618
Is used to represent the word vector of (a),
Figure BDA0002376338810000071
k＝1,2,...,L(qs) For querying qsMiddle j (th) word
Figure BDA0002376338810000072
By using a word vector matrix E ∈ R in a pre-trainingd×|D|The obtained result is searched;
step B34: subjecting the product obtained in step B32
Figure BDA0002376338810000073
And obtained in step B33
Figure BDA0002376338810000074
The GRU network module of the input discriminator network is coded to obtain
Figure BDA0002376338810000075
Subjecting the product obtained in step B32
Figure BDA0002376338810000076
And obtained in step B33
Figure BDA0002376338810000077
The GRU network module of the input discriminator network is coded to obtain
Figure BDA0002376338810000078
And
Figure BDA0002376338810000079
the coding formula of (c) is as follows:
Figure BDA00023763388100000710
Figure BDA00023763388100000711
step B35: subjecting the product obtained in step B34
Figure BDA00023763388100000712
And
Figure BDA00023763388100000713
respectively inputting softmax layers of the discriminator network, and outputting the category probability of the discriminator network considering the discriminator network to belong to the target query predicted by the generator network or the real target query of the user, wherein the calculation formula is as follows:
Figure BDA00023763388100000714
wherein,
Figure BDA00023763388100000715
obtained for step B34
Figure BDA00023763388100000716
Or
Figure BDA00023763388100000717
R represents the probability of belonging to the two categories.
Further, the step B4 specifically includes the following steps:
step B41: regarding the process of generating the query recommendation by the model as an action sequence, regarding the generator network based on the improved VHRED as a strategy, regarding the probability obtained in the step B35 as the reward of the generator network, calculating the loss value:
J(θ)＝E(R-b|θ)
wherein E represents the expected value of the reward, b is a baseline value and is a balance item which enables the training to be stable, and theta is a hyperparameter;
step B42: from the formula of the loss value of step B41, the update gradient is obtained by likelihood approximation:
Figure BDA00023763388100000718
wherein,
Figure BDA00023763388100000719
generating the next word for step B15
Figure BDA00023763388100000720
The probability of (d);
and retraining parameters of the generator network based on the updated gradient, and enabling the target query predicted by the generator network to be closer to the real target query of the user through repeated iterative updating so as to obtain a trained query recommendation deep learning network model.
The invention also provides a query recommendation system adopting the method, which comprises the following steps:
the data collection module is used for collecting all user query log records in the search engine;
the preprocessing module is used for preprocessing the collected user query log record data, extracting queried time characteristic information and text characteristic information and constructing a user query log Training Set (TS);
the network training module is used for training a VHRED model with time characteristics by using the obtained user query log training set, generating target query recommendations by using all queried time characteristic information and text characteristic information in a session, calculating corresponding loss values, training the whole VHRED model with time characteristics by taking a minimum loss value as a target to obtain the trained VHRED model, then enabling a recommendation query generated by the trained VHRED model and a query generated by a real user to pass through a discriminator network to obtain a probability value R as a reward, continuously modifying a learning rate learning _ rate by using the reward R, and controlling a gradient descending direction, so that parameters of a generator network based on the improved VHRED are retrained, and finally the required trained query recommendation deep learning network model is obtained by carrying out repeated iterative updating; and
and the query recommendation module is used for receiving the query sentence input by the user, inputting the query sentence into the trained query recommendation deep learning network model and outputting the matched query recommendation.
Compared with the prior art, the invention has the following beneficial effects: the method and the system generate query recommendation by constructing and training a query recommendation deep learning network model based on the VHRED with the time characteristics and the reinforcement learning, are fast and robust, can accurately know the query intention of a user, generate the query recommendation meeting the needs of the user, and have better practicability and higher application value.
Drawings
Fig. 1 is a flowchart of a method implementation according to an embodiment of the present invention.
Fig. 2 is a schematic structural diagram of a system according to an embodiment of the present invention.
Detailed Description
The invention is described in further detail below with reference to the figures and the embodiments.
The invention provides a query recommendation method based on improved VHRED and reinforcement learning, which comprises the following steps of:
step A: collecting user query log records of a search engine, preprocessing the user query log record data, and constructing a user query log Training Set (TS). The method specifically comprises the following steps:
step A1: collecting user query log records of a search engine to obtain an original query log set; wherein each query log of the search engine is represented by a triplet (u, q, t), u representing the user, q representing the query, and t representing the query time.
Step A2: and dividing the original query log set according to users, and sequencing according to query time to obtain query log subsets of different users.
Step A3: setting a time interval T according to the following rule: query logs with query time intervals larger than T belong to different sessions, queries in different sessions are not related to each other, the last query in the same session is a target query containing a query intention of a user, a query log subset of each user is further divided into a plurality of sessions to obtain a session set of each user, and the session sets of all the users form a user query log training set TS;
one session of one user u in TS is represented as
Figure BDA0002376338810000091
Wherein q isiIndicating the ith query, t, in the sessioniDenotes qiCorresponding to the query time, the session contains ku+1 queries, last query
Figure BDA0002376338810000092
Querying a real target of a user;
to q isiAfter word segmentation and stop word removal, q is addediIs further shown as
Figure BDA0002376338810000093
Denotes qiThe j-th word in (1, 2., L (q) ·i)，L(qi) Represents qiThe number of words of; q. q.siCorresponding query time tiIs denoted by ti＝(xi,yi,zi,di)，xiRepresents hour, yiRepresents minute, ziDenotes second, diIndicating the day of the week.
And B, step B: and training a query recommendation deep learning network model based on VHRED with time characteristics and reinforcement learning by using a user query log Training Set (TS).
The query recommendation deep learning network model comprises a generator network based on a Variable Hierarchical Encoder-Decoder Recurrent neural network VHRED (Variable Hierarchical Recurrent Encoder-Decoder) with time characteristics and a discriminator network based on a Hierarchical Auto Encoder (Hierarchical Auto Encoder), wherein the Hierarchical Auto Encoder encodes words, sentences and paragraphs by multiple layers of GRUs respectively to capture semantic structure information of different levels (word level, sentence level and paragraph level). The step B specifically comprises the following steps:
step B1: and inputting query text and query time pairs in a training set TS of the user query log into the generator network by taking the user session as a unit, and outputting the target query predicted by the generator network. The method specifically comprises the following steps:
step B11: taking user session as unit, and sending the user session to the clientEach query text to query time pair (q) in the conversation except for the target queryi,ti) Coding to obtain a characterization vector
Figure BDA0002376338810000094
If the user session is
Figure BDA0002376338810000095
A query text to query time pair (q) in the user sessioni,ti) Is characterized by a token vector
Figure BDA0002376338810000096
Is shown as
Figure BDA0002376338810000097
i＝1,2,...,ku，
Figure BDA0002376338810000098
To characterize a vector
Figure BDA0002376338810000101
And
Figure BDA0002376338810000102
the connection of (a) to (b),
Figure BDA0002376338810000103
is q isiThe characterization vector of (a) is calculated,
Figure BDA0002376338810000104
is tiThe characterization vector of (2);
wherein q isiThe coding formula of (a) is as follows:
Figure BDA0002376338810000105
wherein GRU represents a gated recurrent neural network,
Figure BDA0002376338810000106
j＝1,2,...,L(qi) Is qiThe j-th word in
Figure BDA0002376338810000107
Is used to represent the word vector of (a),
Figure BDA0002376338810000108
by using a pre-trained word vector matrix E ∈ Rd×|D|Where D represents the dimension of the word vector, | D | is the number of words in the lexicon D,
Figure BDA0002376338810000109
tithe coding formula of (a) is as follows:
Figure BDA00023763388100001010
wherein,
Figure BDA00023763388100001011
connection characterization vector
Figure BDA00023763388100001012
And
Figure BDA00023763388100001013
to obtain
Figure BDA00023763388100001014
Step B12: with the user session as a unit, forming a sequence of the characterization vectors of each query text and query time pair except for the target query in the user session, inputting the sequence into an encoder module of a generator network based on a GRU network for encoding to obtain a user session SuIs characterized by a token vector
Figure BDA00023763388100001015
All query information except the target query in the user session is contained;
wherein S isuThe coding formula of (a) is as follows:
Figure BDA00023763388100001016
step B13: obtained according to step B12
Figure BDA00023763388100001017
Computing user sessions SuLatent variable z ofu；
Firstly, the first step is to
Figure BDA00023763388100001018
The mean value mu is obtained by the feedforward neural network module of the input generator networkuThe formula is as follows:
Figure BDA00023763388100001019
wherein
Figure BDA00023763388100001020
dzIs a hidden variable zuTanh is the hyperbolic tangent function, fFNNIs a feedforward neural network;
the mean value muuInputting a softplus function, and calculating to obtain covariance sigmauThe formula is as follows:
∑u＝softplus(f(μu))
wherein,
Figure BDA00023763388100001021
softplus is an activation function, softplus (x) log (1+ e)x)；
Then, according to the mean value muuSum covariance ∑uObtaining the latent variable z by random sampling calculationuThe formula is as follows:
Figure BDA0002376338810000111
wherein samples is a random number vector,
Figure BDA0002376338810000112
extracting d from a standard normal distributionzA random number, constituting a random number vector samples,
Figure BDA0002376338810000113
is a vector sigmauHadamard product with samples to obtain hidden variables of user session
Figure BDA0002376338810000114
Step B14: subjecting the product obtained in step B12
Figure BDA0002376338810000115
And z obtained in step B13uThe GRU network-based decoder module which is input into the generator network decodes and outputs the target query predicted by the generator network;
first, a token vector of a user session is calculated
Figure BDA0002376338810000116
Initial hidden state of
Figure BDA0002376338810000117
The formula is as follows:
Figure BDA0002376338810000118
wherein, W1Is a weight parameter that is a function of,
Figure BDA0002376338810000119
dhis h0Dimension of (b)0Is a bias term;
firstly, h obtained by the above formula0Decoding is carried out through a GRU network, and the hidden state vector obtained by decoding enters through the GRU networkLine decoding, repetition KtargetSecondary decoding to generate a block containing KtargetWord-by-word target query qsWherein the decoding formula is as follows:
Figure BDA00023763388100001110
each decoding produces the next word
Figure BDA00023763388100001111
The probability of (c) is:
Figure BDA00023763388100001112
Figure BDA00023763388100001113
wherein f is the full junction layer, W2、W3、bprobParameters of the fully connected layer, bprobIn order to be a bias term, the bias term,
Figure BDA00023763388100001114
W3∈Rd×d，bprob∈Rd；
candidate target query q'sScore of(s) (score of q's) For the probability product of decoding, the formula is as follows:
Figure BDA00023763388100001115
wherein
Figure BDA00023763388100001116
To constitute target query q'sThe sequence of words of (a);
taking logarithm of the above formula, the formula is as follows:
Figure BDA0002376338810000121
selecting
Figure BDA0002376338810000122
Highest q'sTarget query q as predicted by a generator networks：
Figure BDA0002376338810000123
Step B2: and calculating the gradient of each parameter in the generator network by using a back propagation method according to the target loss function loss, and updating the parameter by using a random gradient descent method.
Wherein the target loss function loss is defined as follows:
loss＝loss1+loss2
therein, loss1The loss value loss is obtained using KL divergence measure as the difference between the distribution of hidden variables and the unit Gaussian distribution1The calculation is as follows:
Figure BDA0002376338810000124
wherein, muu,ΣuThe mean and covariance obtained in step B13;
loss2target query q predicted for a generator networksAnd a session SuTarget query representing real query intention of user
Figure BDA0002376338810000125
The loss value calculated by the cross entropy loss function is calculated as follows:
Figure BDA0002376338810000128
wherein Cross EntropyLoss is a cross entropy loss function;
and updating the learning rate through a gradient optimization algorithm AdaGrad, and updating model parameters through back propagation iteration so as to minimize a loss function to train the model.
Step B3: inputting the target query predicted in the step B1 and the real target query of the user in the user session into the discriminator network, outputting the category probability, and judging whether the input target query is the target query predicted by the generator network or the real target query of the user according to the category probability. The method specifically comprises the following steps:
step B31: user session S obtained by B11uRemoving user's true target query
Figure BDA0002376338810000126
Outer text of each query qi,i＝1,2,...,kuInputting the data into a discriminator network for coding to obtain qiIs characterized by a token vector
Figure BDA0002376338810000127
i＝1,2,...,ku；
Step B32: q obtained in the step B32iIs characterized by a token vector
Figure BDA0002376338810000131
Constituting a sequence
Figure BDA0002376338810000132
The GRU network module in the input discriminator network is coded to obtain
Figure BDA0002376338810000133
The coding formula is as follows:
Figure BDA0002376338810000134
step B33: respectively conversing users SuUser's true target query in
Figure BDA0002376338810000135
Target query q predicted from step B14sThe GRU network module of the input discriminator network is coded to obtain
Figure BDA0002376338810000136
And q issIs characterized by a token vector
Figure BDA0002376338810000137
And
Figure BDA0002376338810000138
the coding formula is as follows:
Figure BDA0002376338810000139
Figure BDA00023763388100001310
wherein,
Figure BDA00023763388100001311
for querying
Figure BDA00023763388100001312
Middle j (th) word
Figure BDA00023763388100001313
Is used to represent the word vector of (a),
Figure BDA00023763388100001314
k＝1,2,...,L(qs) For querying qsMiddle j (th) word
Figure BDA00023763388100001315
By using a word vector matrix E ∈ R in a pre-trainingd×|D|The obtained result is searched;
step B34: subjecting the product obtained in step B32
Figure BDA00023763388100001316
And obtained in step B33
Figure BDA00023763388100001317
Inputting GRU network module of the discriminator network for coding to obtain
Figure BDA00023763388100001318
Subjecting the product obtained in step B32
Figure BDA00023763388100001319
And obtained in step B33
Figure BDA00023763388100001320
The GRU network module of the input discriminator network is coded to obtain
Figure BDA00023763388100001321
And
Figure BDA00023763388100001322
the coding formula of (a) is as follows:
Figure BDA00023763388100001323
Figure BDA00023763388100001324
step B35: subjecting the product obtained in step B34
Figure BDA00023763388100001325
And
Figure BDA00023763388100001326
respectively inputting softmax layers of the discriminator network, and outputting the category probability of the discriminator network considering the discriminator network to belong to the target query predicted by the generator network or the real target query of the user, wherein the calculation formula is as follows:
Figure BDA00023763388100001327
wherein,
Figure BDA00023763388100001328
obtained for step B34
Figure BDA00023763388100001329
Or
Figure BDA00023763388100001330
R represents the probability of belonging to the two categories.
Step B4: and B3, taking the class probability output by the discriminator network in the step B3 as the reward of the generator network, and performing reinforcement learning training by using a strategy gradient method to maximize the return expectation. The method specifically comprises the following steps:
step B41: regarding the process of generating the query recommendation by the model as an action sequence, regarding the generator network based on the improved VHRED as a strategy, regarding the probability obtained in the step B35 as the reward of the generator network, calculating the loss value:
J(θ)＝E(R-b|θ)
wherein E represents the expected value of the reward, b is a base line value and is a balance item which enables the training to be stable, and theta is a hyperparameter;
step B42: the update gradient is obtained by likelihood approximation from the formula of the loss value of step B41:
Figure BDA0002376338810000141
wherein,
Figure BDA0002376338810000142
generating the next word for step B15
Figure BDA0002376338810000143
The probability of (d);
if the reward R for an action is large, the probability of next generation of the sequence increases, and for sequences with lower reward R, the generation is relatively suppressed, so that a base value b is subtracted, so that the reward R has a positive or negative value.
In brief, the probability R is obtained as an incentive, the learning rate learning _ rate is modified, the gradient descending direction is controlled, the parameters of the generator network are retrained based on the updated gradient, and the target query predicted by the generator network is closer to the real target query of the user through repeated iteration updating, so that the trained query recommendation deep learning network model is obtained.
Step B5: and when the iterative change of the loss value generated by inquiring the recommended deep learning network model is smaller than a set threshold value or reaches the maximum iteration times, terminating the training of the inquiring recommended deep learning network model.
Step C: and the query recommendation system receives the query sentence input by the user, inputs the query sentence into the trained query recommendation deep learning network model and outputs the matched query recommendation.
The invention also provides a query recommendation system adopting the method, as shown in fig. 2, comprising:
the data collection module is used for collecting all user query log records in the search engine;
the preprocessing module is used for preprocessing the collected user query log record data, extracting queried time characteristic information and text characteristic information and constructing a user query log Training Set (TS);
the network training module is used for training a VHRED model with time characteristics by using the obtained user query log training set, generating target query recommendations by using all queried time characteristic information and text characteristic information in a session, calculating corresponding loss values, training the whole VHRED model with time characteristics by taking a minimum loss value as a target to obtain the trained VHRED model, then enabling a recommendation query generated by the trained VHRED model and a query generated by a real user to pass through a discriminator network to obtain a probability value R as a reward, continuously modifying a learning rate learning _ rate by using the reward R, and controlling a gradient descending direction, so that parameters of a generator network based on the improved VHRED are retrained, and finally the required trained query recommendation deep learning network model is obtained by carrying out repeated iterative updating; and
and the query recommendation module is used for receiving the query sentence input by the user, inputting the query sentence into the trained query recommendation deep learning network model and outputting the matched query recommendation.
The above are preferred embodiments of the present invention, and all changes made according to the technical scheme of the present invention that produce functional effects do not exceed the scope of the technical scheme of the present invention belong to the protection scope of the present invention.

Claims (7)

1. A query recommendation method based on improved VHRED and reinforcement learning is characterized by comprising the following steps:
step A: collecting user query log records of a search engine, preprocessing the user query log record data, and constructing a user query log Training Set (TS);
and B: training a query recommendation deep learning network model based on VHRED with time characteristics and reinforcement learning by using a user query log Training Set (TS);
and C: the query recommendation system receives a query sentence input by a user, inputs the query sentence into the trained query recommendation deep learning network model, and outputs a matched query recommendation;
the inquiry recommendation deep learning network model comprises a generator network based on a variable hierarchical encoder-decoder recurrent neural network (VHRED) with time characteristics and a discriminator network based on a hierarchical self-encoder, wherein the hierarchical self-encoder encodes words, sentences and paragraphs by multiple layers of GRUs respectively so as to capture semantic structure information of different levels; the step B specifically comprises the following steps:
step B1: inputting query text and query time pairs in a user query log training set TS into a generator network by taking a user session as a unit, and outputting a target query predicted by the generator network;
step B2: calculating the gradient of each parameter in the generator network by using a back propagation method according to the target loss function loss, and updating the parameters by using a random gradient descent method;
step B3: inputting the target query predicted in the step B1 and the real target query of the user in the user session into a discriminator network, outputting category probability, and judging whether the input target query is the target query predicted by the generator network or the real target query of the user according to the category probability;
step B4: taking the class probability output by the discriminator network in the step B3 as the reward of the generator network, and performing reinforcement learning training by using a strategy gradient method to maximize the return expectation;
step B5: and when the iterative change of the loss value generated by inquiring the recommended deep learning network model is smaller than a set threshold value or reaches the maximum iteration times, terminating the training of the inquiring recommended deep learning network model.
2. The method of claim 1, wherein the step a specifically comprises the following steps:
step A1: collecting user query log records of a search engine to obtain an original query log set; wherein each query log of the search engine is represented by a triplet (u, q, t), u representing a user, q representing a query, and t representing a query time;
step A2: dividing an original query log set according to users, and sequencing according to query time to obtain query log subsets of different users;
step A3: setting a time interval T according to the following rule: query logs with query time intervals larger than T belong to different sessions, queries in different sessions are not related to each other, the last query in the same session is a target query containing a query intention of a user, a query log subset of each user is further divided into a plurality of sessions to obtain a session set of each user, and the session sets of all the users form a user query log training set TS;
one session of one user u in TS is represented as
Figure FDA0003626477990000021
Wherein q isiIndicating the ith query, t, in the sessioniDenotes qiCorresponding to the query time, the session contains ku+1 queries, last query
Figure FDA0003626477990000022
Querying a real target of a user;
to q isiAfter word segmentation and stop word removal, q is addediIs further shown as
Figure FDA0003626477990000023
Represents qiThe j-th word in (1, 2., L (q) ·i)，L(qi) Denotes qiThe number of words of; q. q.siCorresponding query time tiIs denoted by ti＝(xi,yi,zi,di)，xiRepresents hour, yiRepresents minute, ziDenotes second, diIndicating the day of the week.
3. The method for query recommendation based on VHRED and reinforcement learning of claim 1, wherein said step B1 is to input query text and query time pairs in a training set TS of user query logs into a generator network and output target queries predicted by the generator network in units of user sessions, and comprises the following steps:
step B11: taking the user session as a unit, and carrying out the text and query time pair (q) on each query except the target query in the user sessioni,ti) Coding to obtain a characterization vector
Figure FDA0003626477990000024
If the user session is
Figure FDA0003626477990000025
A query text to query time pair (q) in the user sessioni,ti) Is characterized by a token vector
Figure FDA0003626477990000026
Is shown as
Figure FDA0003626477990000027
i＝1,2,...,ku，
Figure FDA0003626477990000028
To characterize a vector
Figure FDA0003626477990000029
And
Figure FDA00036264779900000210
the connection of (a) to (b),
Figure FDA00036264779900000211
is qiThe characterization vector of (a) is determined,
Figure FDA00036264779900000212
is tiThe characterization vector of (2);
wherein q isiThe coding formula of (a) is as follows:
Figure FDA00036264779900000213
wherein GRU represents a gated recurrent neural network,
Figure FDA00036264779900000214
j＝1,2,...,L(qi) Is qiThe j-th word in
Figure FDA00036264779900000215
Is used to represent the word vector of (a),
Figure FDA00036264779900000216
by using a pre-trained word vector matrix E ∈ Rd×|D|Where D represents the dimension of the word vector, | D | is the number of words in the lexicon D,
Figure FDA00036264779900000217
tithe coding formula of (a) is as follows:
Figure FDA0003626477990000031
wherein,
Figure FDA0003626477990000032
concatenated token vectors
Figure FDA0003626477990000033
And
Figure FDA0003626477990000034
to obtain
Figure FDA0003626477990000035
Step B12: with the user session as a unit, forming a sequence of the characterization vectors of each query text and query time pair except for the target query in the user session, inputting the sequence into an encoder module of a generator network based on a GRU network for encoding to obtain a user session SuIs characterized vector of
Figure FDA0003626477990000036
All query information except the target query in the user session is contained;
wherein S isuThe coding formula of (a) is as follows:
Figure FDA0003626477990000037
step B13: obtained according to step B12
Figure FDA0003626477990000038
Computing user sessions SuLatent variable z ofu；
Firstly, the first step is to
Figure FDA0003626477990000039
The mean value mu is obtained by a feedforward neural network module of the input generator networkuThe formula is as follows:
Figure FDA00036264779900000310
wherein
Figure FDA00036264779900000311
dzIs a hidden variable zuTanh is the hyperbolic tangent function, fFNNIs a feedforward neural network;
mean value muuInputting a softplus function, and calculating to obtain covariance sigmauThe formula is as follows:
∑u＝softplus(f(μu))
wherein,
Figure FDA00036264779900000312
softplus is an activation function, softplus (x) log (1+ e)x)；
Then, according to the mean value muuSum covariance ∑uObtaining the latent variable z by random sampling calculationuThe formula is as follows:
Figure FDA00036264779900000313
wherein samples is random number vector, samples∈RdzExtracting d from the normal distributionzA random number constituting a random number vector samples "
Figure FDA00036264779900000314
Is a vector sigmauHadamard product with samples to obtain hidden variables of user session
Figure FDA00036264779900000315
Step B14: subjecting the product obtained in step B12
Figure FDA00036264779900000316
And z obtained in step B13uThe GRU network-based decoder module which is input into the generator network decodes and outputs the target query predicted by the generator network;
first, a token vector of a user session is calculated
Figure FDA0003626477990000041
Initial hidden state of
Figure FDA0003626477990000042
The formula is as follows:
Figure FDA0003626477990000043
wherein, W1Is a weight parameter that is a function of,
Figure FDA0003626477990000044
dhis h0Dimension of (b)0Is a bias term;
firstly, h obtained by the above formula0Decoding through GRU network, decoding the decoded hidden state vector through GRU network, repeating KtargetSecondary decoding to generate a block containing KtargetWord-by-word target query qsWherein the decoding formula is asThe following:
Figure FDA0003626477990000045
each decoding produces the next word
Figure FDA0003626477990000046
The probability of (c) is:
Figure FDA0003626477990000047
Figure FDA0003626477990000048
wherein f is the full junction layer, W2、W3、bprobParameters of the fully connected layer, bprobIn order to be a bias term, the bias term,
Figure FDA0003626477990000049
W3∈Rd×d，bprob∈Rd；
candidate target query q'sScore of(s) (score of q's) For the decoded probability product, the formula is as follows:
Figure FDA00036264779900000410
wherein
Figure FDA00036264779900000411
To constitute target query q'sThe word sequence of (a);
taking the logarithm of the above formula, the formula is as follows:
Figure FDA00036264779900000412
selecting
Figure FDA00036264779900000413
Highest q'sTarget query q as predicted by a generator networks：
Figure FDA00036264779900000414
4. The method for recommending queries based on VHRED and reinforcement learning of claim 3, wherein in said step B2, the objective loss function loss is defined as follows:
loss＝loss1+loss2
among them, loss1The loss value loss is obtained using KL divergence measure as the difference between the distribution of hidden variables and the unit Gaussian distribution1The calculation is as follows:
Figure FDA0003626477990000051
wherein, muu,∑uThe mean and covariance obtained in step B13;
loss2target query q predicted for a generator networksAnd a session SuTarget query representing real query intention of user
Figure FDA0003626477990000052
The loss value calculated by the cross entropy loss function is calculated as follows:
Figure FDA0003626477990000053
wherein Cross EntropyLoss is a cross entropy loss function;
and updating the learning rate by a gradient optimization algorithm AdaGrad, and updating model parameters by using back propagation iteration to train the model by minimizing a loss function.
5. The method of claim 4, wherein the step B3 of inputting the target query predicted in the step B1 and the real target query of the user in the user session into the network of discriminators, outputting a category probability, and determining whether the input target query is the target query predicted by the generator network or the real target query of the user according to the category probability comprises the following steps:
step B31: user session S obtained by B11uRemoving user's true target query
Figure FDA0003626477990000054
Outer text of each query qi,i＝1,2,...,kuInputting the data into a discriminator network for coding to obtain qiIs characterized by a token vector
Figure FDA0003626477990000055
i＝1,2,...,ku；
Step B32: q obtained in the step B31iIs characterized by a token vector
Figure FDA0003626477990000056
Constituting a sequence
Figure FDA0003626477990000057
Inputting GRU network module in the discriminator network for coding to obtain
Figure FDA0003626477990000058
The coding formula is as follows:
Figure FDA0003626477990000059
step B33: respectively conversing users SuUser's true target query in
Figure FDA00036264779900000510
Target query q predicted from step B14sThe GRU network module of the input discriminator network is coded to obtain
Figure FDA00036264779900000511
And q issIs characterized by a token vector
Figure FDA00036264779900000512
And
Figure FDA00036264779900000513
the coding formula is as follows:
Figure FDA0003626477990000061
Figure FDA0003626477990000062
wherein,
Figure FDA0003626477990000063
for querying
Figure FDA0003626477990000064
Middle j (th) word
Figure FDA0003626477990000065
Is used to represent the word vector of (a),
Figure FDA0003626477990000066
k＝1,2,...,L(qs) For querying qsMiddle j (th) word
Figure FDA0003626477990000067
By using a word vector matrix E ∈ R in a pre-trainingd×|D|The obtained result is searched;
step B34: subjecting the product obtained in step B32
Figure FDA0003626477990000068
And obtained in step B33
Figure FDA0003626477990000069
The GRU network module of the input discriminator network is coded to obtain
Figure FDA00036264779900000610
Subjecting the product obtained in step B32
Figure FDA00036264779900000611
And obtained in step B33
Figure FDA00036264779900000612
The GRU network module of the input discriminator network is coded to obtain
Figure FDA00036264779900000613
And
Figure FDA00036264779900000614
the coding formula of (a) is as follows:
Figure FDA00036264779900000615
Figure FDA00036264779900000616
step B35: subjecting the product obtained in step B34
Figure FDA00036264779900000617
And
Figure FDA00036264779900000618
respectively inputting softmax layers of the discriminator network, and outputting the category probability of the discriminator network considering the discriminator network to belong to the target query predicted by the generator network or the real target query of the user, wherein the calculation formula is as follows:
Figure FDA00036264779900000619
wherein,
Figure FDA00036264779900000620
obtained for step B34
Figure FDA00036264779900000621
Or
Figure FDA00036264779900000622
R represents the probability of belonging to the two categories.
6. The method for recommending queries based on VHRED and reinforcement learning of claim 5, wherein said step B4 specifically comprises the following steps:
step B41: regarding the process of generating the query recommendation by the model as an action sequence, regarding the generator network based on the improved VHRED as a strategy, regarding the probability obtained in the step B35 as the reward of the generator network, calculating the loss value:
J(θ)＝E(R-b|θ)
wherein E represents the expected value of the reward, b is a baseline value and is a balance item which enables the training to be stable, and theta is a hyperparameter;
step B42: the update gradient is obtained by likelihood approximation from the formula of the loss value of step B41:
Figure FDA00036264779900000623
wherein,
Figure FDA0003626477990000071
generating the next word for step B14
Figure FDA0003626477990000072
The probability of (d);
and retraining parameters of the generator network based on the updated gradient, and enabling the target query predicted by the generator network to be closer to the real target query of the user through repeated iterative updating so as to obtain a trained query recommendation deep learning network model.
7. A query recommendation system employing the method of any of claims 1-6, comprising:
the data collection module is used for collecting all user query log records in the search engine;
the preprocessing module is used for preprocessing the collected user query log record data, extracting queried time characteristic information and text characteristic information and constructing a user query log Training Set (TS);
the network training module is used for training a VHRED model with time characteristics by using the obtained user query log training set, generating target query recommendations by using all queried time characteristic information and text characteristic information in a session, calculating corresponding loss values, training the whole VHRED model with time characteristics by taking a minimum loss value as a target to obtain the trained VHRED model, then enabling a recommendation query generated by the trained VHRED model and a query generated by a real user to pass through a discriminator network to obtain a probability value R as a reward, continuously modifying a learning rate learning _ rate by using the reward R, and controlling a gradient descending direction, so that parameters of a generator network based on the improved VHRED are retrained, and finally the required trained query recommendation deep learning network model is obtained by carrying out repeated iterative updating;
and the query recommendation module is used for receiving the query sentence input by the user, inputting the query sentence into the trained query recommendation deep learning network model and outputting the matched query recommendation.
CN202010067232.4A 2020-01-20 2020-01-20 Query recommendation method and system based on improved VHRED and reinforcement learning Active CN111274359B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010067232.4A CN111274359B (en) 2020-01-20 2020-01-20 Query recommendation method and system based on improved VHRED and reinforcement learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010067232.4A CN111274359B (en) 2020-01-20 2020-01-20 Query recommendation method and system based on improved VHRED and reinforcement learning

Publications (2)

Publication Number Publication Date
CN111274359A CN111274359A (en) 2020-06-12
CN111274359B true CN111274359B (en) 2022-06-14

Family

ID=70998997

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010067232.4A Active CN111274359B (en) 2020-01-20 2020-01-20 Query recommendation method and system based on improved VHRED and reinforcement learning

Country Status (1)

Country Link
CN (1) CN111274359B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114077783A (en) * 2020-08-18 2022-02-22 中国电信股份有限公司 Interactive simulation method and device for recommendation system based on reinforcement learning
CN113360497B (en) * 2021-05-26 2022-04-05 华中科技大学 Multi-load-oriented automatic recommendation method and system for secondary indexes of cloud database
CN115070753A (en) * 2022-04-28 2022-09-20 同济大学 Multi-target reinforcement learning method based on unsupervised image editing

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102609433A (en) * 2011-12-16 2012-07-25 北京大学 Method and system for recommending query based on user log
CN106557563A (en) * 2016-11-15 2017-04-05 北京百度网讯科技有限公司 Query statement based on artificial intelligence recommends method and device
CN107122469A (en) * 2017-04-28 2017-09-01 中国人民解放军国防科学技术大学 Sort method and device are recommended in inquiry based on semantic similarity and timeliness resistant frequency
CN109145213A (en) * 2018-08-22 2019-01-04 清华大学 Inquiry recommended method and device based on historical information

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170185673A1 (en) * 2015-12-25 2017-06-29 Le Holdings (Beijing) Co., Ltd. Method and Electronic Device for QUERY RECOMMENDATION

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102609433A (en) * 2011-12-16 2012-07-25 北京大学 Method and system for recommending query based on user log
CN106557563A (en) * 2016-11-15 2017-04-05 北京百度网讯科技有限公司 Query statement based on artificial intelligence recommends method and device
CN107122469A (en) * 2017-04-28 2017-09-01 中国人民解放军国防科学技术大学 Sort method and device are recommended in inquiry based on semantic similarity and timeliness resistant frequency
CN109145213A (en) * 2018-08-22 2019-01-04 清华大学 Inquiry recommended method and device based on historical information

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
查询推荐研究综述;张晓娟等;《情报学报》;20190424;第38卷(第4期);全文 *

Also Published As

Publication number Publication date
CN111274359A (en) 2020-06-12

Similar Documents

Publication Publication Date Title
CN109543180B (en) Text emotion analysis method based on attention mechanism
CN110222163B (en) Intelligent question-answering method and system integrating CNN and bidirectional LSTM
Kim et al. Domain attention with an ensemble of experts
Yang et al. Multitask learning and reinforcement learning for personalized dialog generation: An empirical study
Wang Bankruptcy prediction using machine learning
CN111274359B (en) Query recommendation method and system based on improved VHRED and reinforcement learning
CN111626063A (en) Text intention identification method and system based on projection gradient descent and label smoothing
CN112417894B (en) Conversation intention identification method and system based on multi-task learning
CN111782961B (en) Answer recommendation method oriented to machine reading understanding
CN112749274B (en) Chinese text classification method based on attention mechanism and interference word deletion
CN112232087B (en) Specific aspect emotion analysis method of multi-granularity attention model based on Transformer
CN113255366B (en) Aspect-level text emotion analysis method based on heterogeneous graph neural network
US20230169271A1 (en) System and methods for neural topic modeling using topic attention networks
CN112182154A (en) Personalized search model for eliminating keyword ambiguity by utilizing personal word vector
CN112416956A (en) Question classification method based on BERT and independent cyclic neural network
CN109308316B (en) Adaptive dialog generation system based on topic clustering
CN118193683B (en) Text recommendation method and system based on language big model
Moriya et al. Evolution-strategy-based automation of system development for high-performance speech recognition
CN115994224A (en) Phishing URL detection method and system based on pre-training language model
CN111444328A (en) Natural language automatic prediction inference method with interpretation generation
CN116167353A (en) Text semantic similarity measurement method based on twin long-term memory network
CN113342964B (en) Recommendation type determination method and system based on mobile service
Lee et al. A two-level recurrent neural network language model based on the continuous Bag-of-Words model for sentence classification
CN108563639B (en) Mongolian language model based on recurrent neural network
Hilmiaji et al. Identifying Emotion on Indonesian Tweets using Convolutional Neural Networks

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant