CN110866403A - End-to-end conversation state tracking method and system based on convolution cycle entity network - Google Patents

End-to-end conversation state tracking method and system based on convolution cycle entity network Download PDF

Info

Publication number
CN110866403A
CN110866403A CN201810916744.6A CN201810916744A CN110866403A CN 110866403 A CN110866403 A CN 110866403A CN 201810916744 A CN201810916744 A CN 201810916744A CN 110866403 A CN110866403 A CN 110866403A
Authority
CN
China
Prior art keywords
sentence
convolution
matrix
semantic
semantic slot
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.)
Granted
Application number
CN201810916744.6A
Other languages
Chinese (zh)
Other versions
CN110866403B (en
Inventor
颜永红
何峻青
赵学敏
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.)
Institute of Acoustics CAS
Original Assignee
Institute of Acoustics CAS
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 Institute of Acoustics CAS filed Critical Institute of Acoustics CAS
Priority to CN201810916744.6A priority Critical patent/CN110866403B/en
Publication of CN110866403A publication Critical patent/CN110866403A/en
Application granted granted Critical
Publication of CN110866403B publication Critical patent/CN110866403B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Machine Translation (AREA)

Abstract

The invention provides an end-to-end conversation state tracking method and system based on a convolution cycle entity network, comprising the following steps: step 1) representing a dialog as a plurality of sentence matrix sets D ═ S1,...St}，SiI is more than or equal to 1 and less than or equal to t is the ith sentence matrix consisting of a plurality of word vectors; step 2) the matrix set D passes through a trainable convolutional neural network module, and sentence vectors with fixed length are obtained after maximal pooling; step 3) using dynamic memory to encode each sentence vector with fixed length, using the last hidden layer h of dynamic memorytRepresenting the entire conversation; step 4) establishing a layer of slave h for each predefined semantic slottObtaining probability distribution of each semantic groove on each value by a fully-connected neural network of all possible values of the semantic groove;and 5) taking the value of the maximum probability as the prediction result of the semantic slot to obtain the current conversation state of the conversation. The invention can automatically learn the text representation related to the semantic slot, and improves the performance of dialog state tracking.

Description

End-to-end conversation state tracking method and system based on convolution cycle entity network
Technical Field
The invention relates to the field of session state tracking of a session system, in particular to an end-to-end session state tracking method and system based on a convolution cycle entity network.
Background
Dialog state tracking is an important component in task-based dialog systems, whose goal is to maintain and update the user's goals, i.e., the values of the various semantic slots in a particular task, from time to time. For example, in a pre-determined restaurant task, querying a restaurant requires three semantic slots: cuisine, price and restaurant location, then dialog state tracking updates the values of these three semantic slots based on user input at all times in a multi-turn dialog.
Given a user input text and dialog history text, how to represent and update dialog states is a long-standing area of research and has recently been combined with deep learning and neural networks to eliminate manual labor. The current Neural Network-based method mainly includes a Convolutional Neural Network (CNN), a cyclic Neural Network (RNN), a Long Short Term Memory unit (LSTM), a Memory Network (MemNN), and a Neural Belief Tracker (NBT). The first four methods do not improve the network structure aiming at the special task of state tracking, are directly used in the task, and lack pertinence. The last method, NBT, requires preprocessing according to the semantic slots and constructs classifiers for each semantic slot value, and is not applicable to semantic slots with a large number of possible values. In addition, the effect of the end-to-end methods on a standard data set commonly used in the industry, namely DSTC2, is still not ideal, and the highest performance only reaches 73.4% of accuracy.
Disclosure of Invention
The invention aims to solve the problems that the existing method does not improve the network structure aiming at the special task of state tracking, lacks pertinence, is not suitable for semantic slots with a large number of possible values and the end-to-end method still has unsatisfactory effect on the standard data set DSTC2 commonly used in the industry.
In order to achieve the above object, the present invention provides an end-to-end conversation state tracking method based on a convolution cycle entity network, the method comprising:
step 1) representing a dialog as a plurality of sentence matrix sets D ═ S1,...St}，SiI is more than or equal to 1 and less than or equal to t is the ith sentence matrix consisting of a plurality of word vectors;
step 2) the matrix set D passes through a trainable convolutional neural network CNN module, and sentence vectors with fixed length are obtained after maximal pooling;
step 3) using dynamic memory to encode each sentence vector with fixed length, and using the last hidden layer h of dynamic memorytRepresenting the entire conversation;
step 4) establishing a layer of slave h for each predefined semantic slottObtaining probability distribution of each semantic groove on each value by a fully-connected neural network of all possible values of the semantic groove;
and 5) taking the value of the maximum probability as the prediction result of the semantic slot to obtain the current conversation state of the conversation.
As a modification of the above method, the step 1) includes:
step 1-1) cutting dialogue data into t sentences according to each wheel dialogue, wherein the ith sentence, i is more than or equal to 1 and is less than or equal to t sentences contain a plurality of words, each word is represented by a word vector with fixed length, and the ith sentence, i is more than or equal to 1 and is less than or equal to t sentences are represented as a sentence matrix SiFor each sentence matrix SiThe sentence matrix SiThe number of lines is the number of word vectors contained in the sentence, the sentence matrix SiThe number of columns of (a) is the dimension of the word vector;
step 1-2) representing the dialogue data as a plurality of sentence matrix sets D ═ S1,...St}。
As a modification of the above method, the step 2) includes:
step 2-1) for a convolution kernel W of height zmUsing it as sliding step length with 1, in sentence matrix SiSliding from top to bottom, calculating the dot product sum x of two matrixes of the overlapped part in each stepiObtaining a vector X with the length of N-z + 1:
xi＝ReLU(Wm·Si:i+z-1+bm) (1)
X＝[x1,x2,...,xN-z+1](2)
wherein, for dot product operation, Si:i+z-1Represents the ith through i + z-1 th rows of the sentence matrix.]Representing element concatenation, ReLU regular linear operation, bmN is the number of words contained in the sentence for the bias of the corresponding convolution kernel; i is the ith step of convolution kernel sliding; m is the serial number of the convolution kernel;
step 2-2) using maximum pooling for the vector X, taking the maximum value to obtain an element cm；
cm＝max(X) (3)
Step 2-3) performing convolution by using a plurality of different convolution kernels, wherein the height and the width of each convolution kernel are the length of the word vector, and the step 2-1) and the step 2-2) are performed for a plurality of times, so that the maximum value c of the vector X obtained by each convolution is obtainedmAnd (3) splicing to obtain a sentence vector s:
s＝[c1,c2,...,ck](4)
where k is the total number of convolution kernels.
As a modification of the above method, the step 3) includes:
step 3-1) inputting sentence vectors s obtained from each sentence into dynamic memory;
step 3-2) inputting the t sentence vector stDynamic memory of a block j
Figure BDA0001763234350000031
The calculation formula is as follows:
Figure BDA0001763234350000032
Figure BDA0001763234350000033
Figure BDA0001763234350000034
Figure BDA0001763234350000035
wherein,
Figure BDA0001763234350000036
to update the gate, σ is the sigmoid function, wjFor the trainable key vector of each block,
Figure BDA0001763234350000037
for the purpose of the updated candidate state,
Figure BDA0001763234350000038
for any non-linear activation function,
Figure BDA0001763234350000039
a hidden layer of the jth sentence, U, V and W represent trainable matrix parameters; t represents matrix transposition;
step 3-3) hidden layer vectors of all blocks at the moment
Figure BDA00017632343500000310
Splicing to obtain a hidden layer h at the momentt：
Figure BDA00017632343500000311
Step 3-4) taking the dynamic memory hidden layer h of the last sentencetIndicating the round of dialog.
As a modification of the above method, the step 4) includes:
step 4-1) for hidden layer htAll possible values of each semantic slot comprise two external choices of non and Dongcare, and a layer of neural network is established;
step 4-2) using Softmax to carry out normalization to obtain the probability y' of each possible value, wherein the formula is as follows:
y'＝Softmax(Rht) (10)
wherein R is a dynamic memory hidden layer h from the momenttThe parameter matrix mapped to the semantic slot, y' is the probability estimate of all values on the semantic slot.
As a modification of the above method, the step 5) includes:
during training, the cross entropy of the true probability distribution y and the predicted probability distribution y' is used as a loss function loss, and the loss function is minimized to adjust all trainable parameters including convolution kernels in a convolution network; adjusting parameters by using a back propagation algorithm;
Figure BDA00017632343500000312
Figure BDA00017632343500000313
wherein M is the number of predefined semantic slots, i is the ith semantic slot, y'iProbability estimation, y, representing all values on the ith semantic slotiFor all values true probability distribution, V, on the ith semantic slotiIndicates the number of values contained in the ith semantic slot, j indicates the jth value in the semantic slot,
Figure BDA0001763234350000041
respectively representing the probability corresponding to the jth element in the ith semantic slot in the real probability distribution and the estimated probability distribution;
during testing, all trainable parameters are loaded with corresponding values from the trained model; and for each semantic slot, taking the option corresponding to the maximum probability value as a prediction result to obtain a predicted dialogue state.
The invention also provides an end-to-end conversation state tracking system based on a convolution cycle entity network, which comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein the processor realizes the steps of the method when executing the program.
The invention has the advantages that:
1. the method can automatically learn the text representation related to the semantic slot, and can obtain the text representation related to the semantic slot on the semantic representation through the convolutional neural network;
2. in the invention, on the aspect of state tracking, a circulating entity network with blocks is used for coding information related to semantic slots, thereby updating the state and realizing better effects than the commonly used RNN, LSTM and the like;
3. the invention uses less parameters in time complexity and space complexity, and is superior to the existing model;
4. the invention uses the convolution cycle entity network designed aiming at the dialogue tracking task, thereby improving the dialogue state tracking performance.
Drawings
FIG. 1 is a block diagram of the method of the present invention;
FIG. 2 is a schematic diagram of the structure of the method of the present invention.
Detailed Description
The invention is described in detail below with reference to the figures and specific embodiments.
At present, a Recurrent neural Network (Recurrent Entity Network) is proposed, and answers to questions can be effectively tracked on a given story for a question-answering task; the performance is greatly superior to LSTM, MemNN. Therefore, the invention aims at the dialog state tracking task, improves the model and provides a Convolutional Recurrent Entity Network (CREN). The network system comprises 3 major parts: convolutional neural networks, dynamic memory, and semantic slot classifiers. The convolutional neural network is responsible for expressing each sentence by semantic slot correlation, dynamic memory is used for further coding and updating all sentence expressions of the whole dialogue, and a semantic slot classifier is used for carrying out probability estimation on a value of each semantic slot which is defined in advance. The model can automatically learn the text representation related to the semantic slot, and the dynamic memory used can encode the value of the semantic slot by using different blocks (Block), thereby updating the state.
The invention discloses an end-to-end dialogue state tracking method based on a convolution cycle entity network. For example, assume that we define a restaurant domain dialog system with two semantic slots: food, location. Inputting a one-way dialog: { "How can I hellpyou? "find a Chinese resource in the source part of top" }, after passing through the convolution cycling entity network, output { food: Chinese, location: source }.
The structure of the entire convolutional cyclic entity network is shown in fig. 1, with D ═ S for one dialogue1,...StWhere t is the number of sentences, StIs a sentence. For each word of each sentence, the word vector is firstly expressed, and then the whole sentence matrix passes through a trainable CNN module, Max-posing, to obtain a vector with a fixed length. Then, each input sentence vector is coded by using an RNN variant, namely Dynamic Memory (Dynamic Memory), and the last hidden layer h of the Dynamic Memory is usedtRepresenting the entire conversation. Finally, for each semantic slot, a layer of slave h is establishedtAnd obtaining the probability distribution of each value of each semantic slot by a fully connected neural network of all possible values of the semantic slot. And taking the value corresponding to the maximum probability as a prediction result of the semantic slot, so as to obtain the current conversation state of the conversation.
The dynamic memory is divided into different blocks, hidden layers are calculated respectively, and finally the hidden layers are spliced together.
As shown in FIG. 2, each block in the dynamic memory has a respective key vector wi. For a certain block i, for the input sentence vector stFirst, the key and the hidden layer h of the previous timet-1Are counted togetherCalculating the value g of the update gateiAnd candidate hidden layer states
Figure BDA0001763234350000051
Then calculating the hidden layer state of the block, and splicing the hidden layer states of all the blocks to obtain a complete ht. In the figure fθRepresenting an update formula.
In the above technical solution, the method specifically includes:
step S1) cuts the dialogue data per wheel, expressed in the form from the dialogue start to the current dialogue statement set D and the corresponding dialogue state Slot. For each wheel set the sentence set D ═ { u ═ u }1,u2,…,uiExpressing each word in each sentence by a word vector with fixed length, and then expressing a dialog as a plurality of sentence matrix sets D ═ S1,S2,…,SiS, each sentence matrixiIs the set maximum sentence length, which is the number of word vectors contained in the sentence, SiIs the dimension of the word vector;
step S2) matrix S for each sentencetThe process through a convolutional neural network is as follows: for a convolution kernel Wm with height z, which is used to slide from top to bottom in the whole matrix with 1 as sliding step, the dot product of two matrixes in the overlapped part and the activated value x are calculated in each stepi：
xi＝ReLU(Wm·Si:i+z-1+bm) (1)
Finally, a vector X with the length of N-z +1 is obtained, wherein N is the number of words contained in the sentence:
X＝[x1,x2,...,xN-z+1](2)
then using maximum pooling, taking the maximum value to obtain an element cm：
cm＝max(X) (3)
Performing convolution by using a plurality of convolution kernels with different heights, wherein the widths of the convolution kernels are all the length of a word vector, and each convolution is obtainedMaximum value c of vector X ofmAnd (3) splicing to obtain a sentence vector s:
s＝[c1,c2,...,ck](4)
wherein, is a dot product operation.]Representing element concatenation, ReLU representing regular linear Unit (regularizer Unit), k being the total number of convolution kernels, m being the mth convolution, bmAn offset for the corresponding convolution kernel;
step S3) the sentence vector S obtained from each sentence is input into the dynamic memory, and the dynamic memory hidden layer of the last sentence is taken as the representation of the dialog in the round.
Hidden layer for dynamically memorizing a certain block j for inputting t-th sentence vector
Figure BDA0001763234350000061
The calculation formula is as follows:
Figure BDA0001763234350000062
Figure BDA0001763234350000063
Figure BDA0001763234350000064
Figure BDA0001763234350000065
wherein,
Figure BDA0001763234350000066
to update the gate, σ is the sigmoid function, wjA key vector for each tile (trainable);
Figure BDA0001763234350000067
for any non-linear activation function (here ReLU is used),
Figure BDA0001763234350000068
for the purpose of the updated candidate state,
Figure BDA0001763234350000069
a hidden layer of the jth sentence, U, V and W represent trainable matrix parameters; t represents matrix transposition;
then, the hidden vector quantity of all blocks at the moment is measured
Figure BDA00017632343500000610
Splicing to obtain a hidden layer h at the momentt：
Figure BDA00017632343500000611
Step S4), for all possible values (including two external choices of None and Dontcare) of each semantic slot, a layer of neural network is established, and then the probability of each possible value is obtained by using Softmax for normalization, and the formula is as follows:
y'＝Softmax(Rht) (10)
wherein R is the hidden layer mapping h of dynamic memory from the momenttThe parameter matrix, y', that hits the semantic slot is the probability estimate of all values on the semantic slot.
Step S5) during training, the cross entropy of the true probability distribution y and the predicted probability distribution y' is used as a loss function loss, minimizing the loss function to adjust all trainable parameters, including the convolution kernels in the convolutional network:
Figure BDA0001763234350000071
Figure BDA0001763234350000072
wherein M is the number of predefined semantic slots, i is the ith semantic slot, y'iProbability estimation, y, representing all values on the ith semantic slotiFor all values true probability distribution, V, on the ith semantic slotiIndicating the content of the ith semantic slotThe number of values, j represents the jth value in the semantic slot,
Figure BDA0001763234350000073
and
Figure BDA0001763234350000074
respectively representing the probability corresponding to the jth element in the ith semantic slot in the true probability distribution and the estimated probability distribution.
During testing, all trainable parameters are loaded with corresponding values from the trained model. And for each semantic slot, taking the option corresponding to the maximum probability value as a prediction result to obtain a predicted dialogue state.
The answer generation method of the invention not only can effectively control the content of the generated answer, but also improves the quality of the answer.
Finally, it should be noted that the above embodiments are only used for illustrating the technical solutions of the present invention and are not limited. Although the present invention has been described in detail with reference to the embodiments, it will be understood by those skilled in the art that various changes may be made and equivalents may be substituted without departing from the spirit and scope of the invention as defined in the appended claims.

Claims (7)

1. An end-to-end conversation state tracking method based on a convolution cycle entity network comprises the following steps:
step 1) representing a dialog as a plurality of sentence matrix sets D ═ S1,...St}，SiI is more than or equal to 1 and less than or equal to t is the ith sentence matrix consisting of a plurality of word vectors;
step 2) the matrix set D passes through a trainable convolutional neural network CNN module, and sentence vectors with fixed length are obtained after maximal pooling;
step 3) using dynamic memory to encode each sentence vector with fixed length, and using the last hidden layer h of dynamic memorytRepresenting the entire conversation;
step 4) establishing a layer of slave h for each predefined semantic slottObtaining probability distribution of each semantic groove on each value by a fully-connected neural network of all possible values of the semantic groove;
and 5) taking the value of the maximum probability as the prediction result of the semantic slot to obtain the current conversation state of the conversation.
2. The method for tracking the end-to-end conversation state based on the convolution cyclic entity network as claimed in claim 1, wherein the step 1) comprises:
step 1-1) cutting dialogue data into t sentences according to each wheel dialogue, wherein the ith sentence, i is more than or equal to 1 and is less than or equal to t sentences contain a plurality of words, each word is represented by a word vector with fixed length, and the ith sentence, i is more than or equal to 1 and is less than or equal to t sentences are represented as a sentence matrix SiFor each sentence matrix SiThe sentence matrix SiThe number of lines is the number of word vectors contained in the sentence, the sentence matrix SiThe number of columns of (a) is the dimension of the word vector;
step 1-2) representing the dialogue data as a plurality of sentence matrix sets D ═ S1,...St}。
3. The method for tracking the end-to-end conversation state based on the convolution cyclic entity network as claimed in claim 2, wherein the step 2) comprises:
step 2-1) for a convolution kernel W of height zmUsing it as sliding step length with 1, in sentence matrix SiSliding from top to bottom, calculating the dot product sum x of two matrixes of the overlapped part in each stepiObtaining a vector X with the length of N-z + 1:
xi＝ReLU(Wm·Si:i+z-1+bm) (1)
X＝[x1,x2,...,xN-z+1](2)
wherein, for dot product operation, Si:i+z-1Represents the ith through i + z-1 th rows of the sentence matrix.]Representing element concatenation, ReLU regular linear operation, bmFor offsets corresponding to convolution kernels, N is included in the sentenceThe number of words; i is the ith step of convolution kernel sliding; m is the serial number of the convolution kernel;
step 2-2) using maximum pooling for the vector X, taking the maximum value to obtain an element cm；
cm＝ma1x(X) (3)
Step 2-3) performing convolution by using a plurality of different convolution kernels, wherein the height and the width of each convolution kernel are the length of the word vector, and the step 2-1) and the step 2-2) are performed for a plurality of times, so that the maximum value c of the vector X obtained by each convolution is obtainedmAnd (3) splicing to obtain a sentence vector s:
s＝[c1,c2,...,ck](4)
where k is the total number of convolution kernels.
4. The method for tracking the end-to-end conversation state based on the convolution cyclic entity network as claimed in claim 3, wherein the step 3) comprises:
step 3-1) inputting sentence vectors s obtained from each sentence into dynamic memory;
step 3-2) inputting the t sentence vector stDynamic memory of a block j
Figure FDA0001763234340000021
The calculation formula is as follows:
Figure FDA0001763234340000022
Figure FDA0001763234340000023
Figure FDA0001763234340000024
Figure FDA0001763234340000025
wherein,
Figure FDA0001763234340000026
to update the gate, σ is the sigmoid function, wjFor the trainable key vector of each block,
Figure FDA0001763234340000027
for the purpose of the updated candidate state,
Figure FDA0001763234340000028
for any non-linear activation function,
Figure FDA0001763234340000029
a hidden layer of the jth sentence, U, V and W represent trainable matrix parameters; t represents matrix transposition;
step 3-3) hidden layer vectors of all blocks at the moment
Figure FDA00017632343400000210
Splicing to obtain a hidden layer h at the momentt：
Figure FDA00017632343400000211
Step 3-4) taking the dynamic memory hidden layer h of the last sentencetIndicating the round of dialog.
5. The method for tracking end-to-end conversation state based on convolution cycle entity network as claimed in claim 4, wherein said step 4) includes:
step 4-1) for hidden layer htAll possible values of each semantic slot comprise two external choices of non and Dongcare, and a layer of neural network is established;
step 4-2) using Softmax to carry out normalization to obtain the probability y' of each possible value, wherein the formula is as follows:
y'＝Softmax(Rht) (10)
wherein R is a dynamic memory hidden layer h from the momenttThe parameter matrix mapped to the semantic slot, y' is the probability estimate of all values on the semantic slot.
6. The method for tracking end-to-end conversation state based on convolution cycle entity network as claimed in claim 5, wherein said step 5) includes:
during training, the cross entropy of the true probability distribution y and the predicted probability distribution y' is used as a loss function loss, and the loss function is minimized to adjust all trainable parameters including convolution kernels in a convolution network; adjusting parameters by using a back propagation algorithm;
Figure FDA0001763234340000031
Figure FDA0001763234340000032
wherein M is the number of predefined semantic slots, i is the ith semantic slot, y'iProbability estimation, y, representing all values on the ith semantic slotiFor all values true probability distribution, V, on the ith semantic slotiIndicates the number of values contained in the ith semantic slot, j indicates the jth value in the semantic slot,
Figure FDA0001763234340000033
respectively representing the probability corresponding to the jth element in the ith semantic slot in the real probability distribution and the estimated probability distribution;
during testing, all trainable parameters are loaded with corresponding values from the trained model; and for each semantic slot, taking the option corresponding to the maximum probability value as a prediction result to obtain a predicted dialogue state.
7. A system for tracking end-to-end conversation state based on a convolutional loop entity network, comprising a memory, a processor and a computer program stored on the memory and operable on the processor, wherein the processor executes the program to implement the steps of the method according to any one of claims 1 to 6.
CN201810916744.6A 2018-08-13 2018-08-13 End-to-end conversation state tracking method and system based on convolution cycle entity network Active CN110866403B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810916744.6A CN110866403B (en) 2018-08-13 2018-08-13 End-to-end conversation state tracking method and system based on convolution cycle entity network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810916744.6A CN110866403B (en) 2018-08-13 2018-08-13 End-to-end conversation state tracking method and system based on convolution cycle entity network

Publications (2)

Publication Number Publication Date
CN110866403A true CN110866403A (en) 2020-03-06
CN110866403B CN110866403B (en) 2021-06-08

Family

ID=69650941

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810916744.6A Active CN110866403B (en) 2018-08-13 2018-08-13 End-to-end conversation state tracking method and system based on convolution cycle entity network

Country Status (1)

Country Link
CN (1) CN110866403B (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112287197A (en) * 2020-09-23 2021-01-29 昆明理工大学 Method for detecting sarcasm of case-related microblog comments described by dynamic memory cases
CN114996479A (en) * 2022-06-21 2022-09-02 中国科学院声学研究所 Dialog state tracking method and system based on enhancement technology
CN111462749B (en) * 2020-03-20 2023-07-21 北京邮电大学 End-to-end dialogue system and method based on dialogue state guidance and knowledge base retrieval

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105788593A (en) * 2016-02-29 2016-07-20 中国科学院声学研究所 Method and system for generating dialogue strategy
CN105845137A (en) * 2016-03-18 2016-08-10 中国科学院声学研究所 Voice communication management system
CN106126596A (en) * 2016-06-20 2016-11-16 中国科学院自动化研究所 A kind of answering method based on stratification memory network
US20180137854A1 (en) * 2016-11-14 2018-05-17 Xerox Corporation Machine reading method for dialog state tracking
CN108282587A (en) * 2018-01-19 2018-07-13 重庆邮电大学 Mobile customer service dialogue management method under being oriented to strategy based on status tracking

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105788593A (en) * 2016-02-29 2016-07-20 中国科学院声学研究所 Method and system for generating dialogue strategy
CN105845137A (en) * 2016-03-18 2016-08-10 中国科学院声学研究所 Voice communication management system
CN106126596A (en) * 2016-06-20 2016-11-16 中国科学院自动化研究所 A kind of answering method based on stratification memory network
US20180137854A1 (en) * 2016-11-14 2018-05-17 Xerox Corporation Machine reading method for dialog state tracking
CN108282587A (en) * 2018-01-19 2018-07-13 重庆邮电大学 Mobile customer service dialogue management method under being oriented to strategy based on status tracking

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
TAKASHI USHIO 等: "RECURRENT CONVOLUTIONAL NEURAL NETWORKS FOR STRUCTURED SPEECH ACT TAGGING", 《2016 IEEE SPOKEN LANGUAGE TECHNOLOGY WORKSHOP》 *
任航 等: "口语对话状态追踪的研究", 《网络新媒体技术》 *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111462749B (en) * 2020-03-20 2023-07-21 北京邮电大学 End-to-end dialogue system and method based on dialogue state guidance and knowledge base retrieval
CN112287197A (en) * 2020-09-23 2021-01-29 昆明理工大学 Method for detecting sarcasm of case-related microblog comments described by dynamic memory cases
CN112287197B (en) * 2020-09-23 2022-07-19 昆明理工大学 Method for detecting sarcasm of case-related microblog comments described by dynamic memory cases
CN114996479A (en) * 2022-06-21 2022-09-02 中国科学院声学研究所 Dialog state tracking method and system based on enhancement technology
CN114996479B (en) * 2022-06-21 2024-08-09 中国科学院声学研究所 Dialogue state tracking method and system based on enhancement technology

Also Published As

Publication number Publication date
CN110866403B (en) 2021-06-08

Similar Documents

Publication Publication Date Title
CN111538908B (en) Search ranking method and device, computer equipment and storage medium
US11487950B2 (en) Autonomous evolution intelligent dialogue method, system, and device based on a game with a physical environment
EP3602413B1 (en) Projection neural networks
CN112686058B (en) BERT embedded speech translation model training method and system, and speech translation method and equipment
KR102483643B1 (en) Method and apparatus for training model and for recognizing bawed on the model
CN109710915B (en) Method and device for generating repeated statement
WO2019091020A1 (en) Weight data storage method, and neural network processor based on method
CN110866403B (en) End-to-end conversation state tracking method and system based on convolution cycle entity network
CN111737426B (en) Method for training question-answering model, computer equipment and readable storage medium
US11580406B2 (en) Weight initialization method and apparatus for stable learning of deep learning model using activation function
CN111159419B (en) Knowledge tracking data processing method, system and storage medium based on graph convolution
CN112116090A (en) Neural network structure searching method and device, computer equipment and storage medium
KR20190041819A (en) Apparatus and method for convolution operation of convolution neural network
CN113657595B (en) Neural network accelerator based on neural network real-time pruning
CN110297885B (en) Method, device and equipment for generating real-time event abstract and storage medium
CN113221530B (en) Text similarity matching method and device, computer equipment and storage medium
US20220383119A1 (en) Granular neural network architecture search over low-level primitives
CN109508784A (en) A kind of design method of neural network activation primitive
CN116775843A (en) Question-answer pair evaluation data generation method, question-answer pair evaluation data generation device, computer equipment and storage medium
CN110837567A (en) Method and system for embedding knowledge graph
CN112131261A (en) Community query method and device based on community network and computer equipment
CN116136870A (en) Intelligent social conversation method and conversation system based on enhanced entity representation
CN117828049A (en) Data processing method and related device
CN112000788A (en) Data processing method and device and computer readable storage medium
CN114861671A (en) Model training method and device, computer equipment and storage medium

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