CN111400461A - Intelligent customer service problem matching method and device - Google Patents
Intelligent customer service problem matching method and device Download PDFInfo
- Publication number
- CN111400461A CN111400461A CN201910000948.XA CN201910000948A CN111400461A CN 111400461 A CN111400461 A CN 111400461A CN 201910000948 A CN201910000948 A CN 201910000948A CN 111400461 A CN111400461 A CN 111400461A
- Authority
- CN
- China
- Prior art keywords
- sentence
- vector
- sentence vector
- vectors
- layer
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 54
- 239000013598 vector Substances 0.000 claims description 359
- 239000011159 matrix material Substances 0.000 claims description 51
- 230000002457 bidirectional effect Effects 0.000 claims description 30
- 238000012545 processing Methods 0.000 claims description 24
- 238000004574 scanning tunneling microscopy Methods 0.000 claims description 24
- 230000006870 function Effects 0.000 claims description 19
- 230000003993 interaction Effects 0.000 claims description 18
- 238000004590 computer program Methods 0.000 claims description 17
- 238000013507 mapping Methods 0.000 claims description 14
- 238000004364 calculation method Methods 0.000 claims description 11
- 230000004927 fusion Effects 0.000 claims description 11
- 230000002787 reinforcement Effects 0.000 claims description 11
- 238000003860 storage Methods 0.000 claims description 11
- 238000012549 training Methods 0.000 claims description 10
- 238000009826 distribution Methods 0.000 claims description 7
- 230000009467 reduction Effects 0.000 claims description 7
- 238000005070 sampling Methods 0.000 claims description 7
- 230000003190 augmentative effect Effects 0.000 claims description 6
- 238000004422 calculation algorithm Methods 0.000 description 16
- 230000014509 gene expression Effects 0.000 description 13
- 238000010586 diagram Methods 0.000 description 12
- 230000007246 mechanism Effects 0.000 description 6
- 230000008569 process Effects 0.000 description 6
- 238000004458 analytical method Methods 0.000 description 5
- 238000013528 artificial neural network Methods 0.000 description 4
- 238000011161 development Methods 0.000 description 4
- 230000018109 developmental process Effects 0.000 description 4
- 239000000047 product Substances 0.000 description 4
- 230000009471 action Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 3
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000004048 modification Effects 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 235000017166 Bambusa arundinacea Nutrition 0.000 description 1
- 235000017491 Bambusa tulda Nutrition 0.000 description 1
- 241001330002 Bambuseae Species 0.000 description 1
- 206010063385 Intellectualisation Diseases 0.000 description 1
- 235000015334 Phyllostachys viridis Nutrition 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 230000009118 appropriate response Effects 0.000 description 1
- 238000003491 array Methods 0.000 description 1
- 239000011425 bamboo Substances 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000008451 emotion Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000035784 germination Effects 0.000 description 1
- 235000012907 honey Nutrition 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 230000005012 migration Effects 0.000 description 1
- 238000013508 migration Methods 0.000 description 1
- 230000001617 migratory effect Effects 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000004091 panning Methods 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 239000013589 supplement Substances 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
Images
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides an intelligent customer service problem matching method and device, and belongs to the technical field of intelligent customer service.
Description
Technical Field
The invention relates to the technical field of intelligent customer service, in particular to an intelligent customer service problem matching method and device.
Background
With the germination of a new technological revolution, the formation of big data, the innovation of theoretical algorithm, the improvement of computing power and the evolution of network facilities, the development of artificial intelligence enters a new stage. The falling of a large number of deep learning algorithms enables the field of artificial intelligence to start large-scale development of industrial chain type. Intellectualization has become an important direction in the development of technology and industry. A large amount of intelligent services such as automatic driving, intelligent medical treatment, intelligent manufacturing, intelligent customer service robots and the like realize industrial development and begin to play an important role in the life of people.
Along with the popularization of intelligent services, the acceptance degree of people on intelligent interaction is higher and higher, the adaptability to an automatic process is stronger and stronger, and the market of intelligent product popularization, intelligent product sale and customer service from an internet channel becomes considerable. A large number of intelligent customer service robots emerge like bamboo shoots in spring after rain. The intelligent customer service concept is deep in mind no matter based on the 'small honey' of a commodity after-sale panning robot or the 'Wang young' of a knowledge search dog robot or even based on the 'small ice' of a micro-soft robot which is fond and chatted by a user. The intelligent customer service robots carry out algorithm design and iteration based on respective fields and specific functions, and the method comprises the steps of question matching, emotion analysis, automatic answer generation and the like, and screening of appropriate responses is carried out based on categories facing to customer groups and product self-positioning.
Compared with the fields of finance, e-commerce and the like, the intelligent customer service in the telecommunication field covers a series of functions of telecommunication service popularization, consultation, customization, sale, customer group maintenance and the like. Due to strong functional guidance, the intellectual question and answer search library of the intelligent customer service product in the telecommunication field has the characteristics of strong domain, standard speech and the like. The existing customer service system mechanism based on the intellectual question-answering maintains a candidate question-answering base which comprises well-defined questions and corresponding answers, when a user inputs a conversation question, the system obtains a sentence vector through a method based on semantic analysis, a recurrent neural network or a convolutional neural network, compares the similarity degree of the conversation question and the candidate question through calculating cosine similarity, and selects an answer corresponding to the best matched candidate sentence as a reply of the customer service system to the user question.
The existing intelligent customer service technical scheme based on the intellectual search has the following disadvantages:
1) the candidate question-answer base data is fixed and the mode is fixed. In the existing mode, the search is realized based on the question and answer data which is initially combed by the cold start of the system, and a person who specially sorts the data is arranged to regularly organize new question and answer data according to needs, so that the manual data arrangement mode consumes manpower, the data in the question and answer library is modeled, the method cannot adapt to a large number of flexible question and answer modes of users, and the system accuracy rate is gradually reduced under the condition that the user usage amount is continuously increased;
2) the data volume of the candidate question-answer base based on manual expansion is limited, algorithm training based on big data cannot be realized, and optimization and index improvement of a problem matching algorithm are limited. The continuous increase of daily activity of the system generates a large amount of effective user question and answer logs, and the data cannot be used for expanding and improving the quantity and quality of candidate question and answer libraries, so that the data is very waste;
3) most of the existing problem matching algorithms are based on keyword matching, and few of the existing problem matching algorithms are based on simple convolutional neural network or cyclic neural network models, so that the importance of the precision of sentence vectors in problem matching is ignored, and the accuracy of the algorithm models needs to be improved.
Disclosure of Invention
The technical problem to be solved by the invention is to provide an intelligent customer service problem matching method and device, which can improve the accuracy of the search result of an intelligent customer service system.
To solve the above technical problem, embodiments of the present invention provide the following technical solutions:
the embodiment of the invention provides an intelligent customer service problem matching method, which comprises the following steps:
inputting initial word vectors of sentences 1 and sentences 2 in the candidate question-answering library into a pre-trained reinforced self-attention model to respectively obtain sentence vectors 1 and sentence vectors 2;
inputting sentences 1 and sentences 2 in the candidate question-answering library into a multilayer bidirectional Bi L STM model based on attention interaction to respectively obtain sentence vectors 3 and sentence vectors 4;
obtaining association scores between words of the sentences 1 and 2 and association scores between the words of the sentences 2 and 1 according to the sentence vectors 1, 2, 3 and 4, inputting the association scores into a BI L STM model to obtain final sentence vectors 1 and 2 of the sentences 1 and 2 respectively;
obtaining the similarity between sentence 1 and sentence 2 from the final sentence vector 1 and the final sentence vector 2 and the sentence vector 1 and the sentence vector 2;
and determining answers matched with the user questions according to the similarity between the sentences.
Further, still include:
obtaining user feedback;
and storing the questions and answers fed back by the user as satisfactory questions and answers in the candidate question and answer library.
Further, still include:
dividing the candidate question-answer library into at least one standard answer pool, wherein each standard answer pool contains a standard question corresponding to an answer and a plurality of user questions;
and taking each standard answer pool as a category, and performing multi-classification learning through a self-attention model based on reinforcement learning.
Further, the multi-classification learning through the self-attention model based on reinforcement learning includes:
dividing the user question sentences in the candidate question-answering library into two parts, respectively carrying out word embedding mapping on the first half section and the second half section of the sentence, and respectively generating a sentence vector A and a sentence vector B through the enhanced self-attention model;
and connecting sentence vectors A, B, A + B and A ⊙ B, and optimizing a multi-class cross entropy objective function by using a standard answer pool to which the user question sentence belongs as a label through a full connection layer and softmax.
Further, outputting a sentence vector using the augmented self-attention model comprises:
inputting a sentence vector x ═ x subjected to embedding mapping1,...,xn]；
Respectively obtaining two vectors with the length of n and composed of 1 and 0 through preset probability distribution sampling
Generating a mask matrix:
the score f (x) between the elements of the vector xi,xj) Formed matrix and mask matrix Mij rssBy element addition, a self-attention score f is obtainedrss(xi,xj) For each xjThe softmax operation is carried out and,
to obtain
Deriving a context vector
Combining s and x through a fusion gate to obtain a final sentence vector u;
wherein F is sigmoid (W)(f)[x；s]+b(f)),u＝F⊙x+(1-F)⊙s。
Furthermore, the multilayer bidirectional Bi L STM model consists of 4 layers of Bi L STMs combined with attention attentions, the input of each layer of Bi L STM is formed by connecting the output of the previous layer of hidden layer, the output of the previous layer of attention model and the input of the previous layer, and the hidden layer input of each layer consists of the current input and the hidden layer output of the current layer at the previous moment.
Further, the step of inputting sentences 1 and sentences 2 in the candidate question-answering library into the attention interaction-based multi-layer bidirectional Bi L STM model to obtain sentence vectors 3 and sentence vectors 4 respectively comprises:
inputting word vectors, part-of-speech POS vectors, named entity recognition NER vectors and accurate matching EM vectors of the sentence 1 and the sentence 2 in the candidate question-answering library into the multilayer bidirectional Bi L STM model to obtain a correlation matrix of each word in the sentence 1 and the sentence 2;
and inputting the obtained incidence matrix into a self-coding layer, performing preset feature dimension reduction, and obtaining a sentence vector 3 and a sentence vector 4 through a max-posing layer.
Further, the obtaining of association scores between words of the sentence 1 and the sentence 2 and association scores between the words of the sentence 2 and the sentences 1 according to the sentence vector 1, the sentence vector 2, the sentence vector 3 and the sentence vector 4, and the inputting of the association scores into a BI L STM model to obtain a final sentence vector 1 and a final sentence vector 2 of the sentence 1 and the sentence 2 respectively comprises:
connecting sentence vector 1, sentence vector 2, sentence vector 3 and sentence vector 4 to obtain a similarity matrix;
calculating the association score generated by each word in sentence 1 to each word in sentence 2 according to the similarity matrix;
and inputting the association score into a layer of Bi L STM model to obtain a final sentence vector 1 and a final sentence vector 2 of the sentence 1 and the sentence 2.
Further, the obtaining the similarity between sentence 1 and sentence 2 by using the final sentence vector 1 and final sentence vector 2 and the sentence vector 1 and sentence vector 2 includes:
connecting the final sentence vector 1 and the final sentence vector 2 with the sentence vector 1 and the sentence vector 2 to form a vector [ final sentence vector 1, final sentence vector 2, final sentence vector 1 ⊙, final sentence vector 2, sentence vector 1 ⊙ sentence vector 2 ];
the similarity between sentence 1 and sentence 2 is obtained through the full-connected layer and softmax.
The embodiment of the invention also provides an intelligent customer service problem matching device, which comprises:
the first processing module is used for inputting initial word vectors of sentences 1 and sentences 2 in the candidate question-answering library into a pre-trained reinforced self-attention model to respectively obtain sentence vectors 1 and sentence vectors 2;
the second processing module is used for inputting sentences 1 and sentences 2 in the candidate question-answering library into a multilayer bidirectional Bi L STM model based on attention interaction to respectively obtain sentence vectors 3 and sentence vectors 4;
the third processing module is used for obtaining association scores between words of the sentence 1 and the sentence 2 and association scores between the words of the sentence 2 and the sentences 1 according to the sentence vector 1, the sentence vector 2, the sentence vector 3 and the sentence vector 4, inputting the association scores into a BI L STM model to respectively obtain a final sentence vector 1 and a final sentence vector 2 of the sentence 1 and the sentence 2;
the similarity calculation module is used for obtaining the similarity between the sentence vector 1 and the sentence vector 2 from the final sentence vector 1 and the final sentence vector 2 and the sentence vector 1 and the sentence vector 2;
and the matching module is used for determining answers matched with the user questions according to the similarity between the sentences.
Further, still include:
the storage module is used for acquiring user feedback; and storing the questions and answers fed back by the user as satisfactory questions and answers in the candidate question and answer library.
Further, still include:
the training module is used for dividing the candidate question-answer library into at least one standard answer pool, and each standard answer pool contains standard questions corresponding to answers and a plurality of user questions; and taking each standard answer pool as a category, and performing multi-classification learning through a self-attention model based on reinforcement learning.
Further, the training module is specifically configured to:
dividing the user question sentences in the candidate question-answering library into two parts, respectively carrying out word embedding mapping on the first half section and the second half section of the sentence, and respectively generating a sentence vector A and a sentence vector B through the enhanced self-attention model;
and connecting sentence vectors A, B, A + B and A ⊙ B, and optimizing a multi-class cross entropy objective function by using a standard answer pool to which the user question sentence belongs as a label through a full connection layer and softmax.
Further, the first processing module is specifically configured to output a sentence vector using the augmented self-attention model by:
inputting a sentence vector x ═ x subjected to embedding mapping1,...,xn]；
Respectively obtaining two vectors with the length of n and composed of 1 and 0 through preset probability distribution sampling
Generating a mask matrix:
the score f (x) between the elements of the vector xi,xj) Formed matrix and mask matrix Mij rssBy element addition, a self-attention score f is obtainedrss(xi,xj) For each xjThe softmax operation is carried out and,
to obtain
Deriving a context vector
Combining s and x through a fusion gate to obtain a final sentence vector u;
wherein F is sigmoid (W)(f)[x；s]+b(f)),u＝F⊙x+(1-F)⊙s。
Furthermore, the multilayer bidirectional Bi L STM model consists of 4 layers of Bi L STMs combined with attention attentions, the input of each layer of Bi L STM is formed by connecting the output of the previous layer of hidden layer, the output of the previous layer of attention model and the input of the previous layer, and the hidden layer input of each layer consists of the current input and the hidden layer output of the current layer at the previous moment.
The second processing module is specifically configured to input word vectors, part-of-speech tagging POS vectors, named entity recognition NER vectors, and exact matching EM vectors of sentences 1 and 2 in the candidate question-answering library into the multi-layer bidirectional Bi L STM model to obtain an incidence matrix of each word in the sentence 1 and the sentence 2, input the obtained incidence matrix into a self-coding layer, perform preset feature dimension reduction, and obtain a sentence vector 3 and a sentence vector 4 through a max-posing layer.
The third processing module is specifically configured to link the sentence vector 1, the sentence vector 2, the sentence vector 3, and the sentence vector 4 to obtain a similarity matrix, calculate a relevance score generated by each word in the sentence 1 for each word in the sentence 2 according to the similarity matrix, and input the relevance score into a Bi L STM model to obtain a final sentence vector 1 and a final sentence vector 2 of the sentence 1 and the sentence 2.
Further, the similarity calculation module is specifically configured to connect the final sentence vector 1 and the final sentence vector 2 with the sentence vector 1 and the sentence vector 2 to form vectors [ the final sentence vector 1; the final sentence vector 2; the final sentence vector 1 ⊙; the final sentence vector 2; the sentence vector 1 ⊙; and the sentence vector 2], and obtain the similarity between the sentences 1 and 2 through a full connection layer and softmax.
The embodiment of the invention also provides an intelligent customer service problem matching device, which comprises: a memory, a processor and a computer program stored on the memory and executable on the processor, the computer program when executed by the processor implementing the steps in the intelligent customer service problem matching method as described above.
An embodiment of the present invention further provides a computer-readable storage medium, where a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the steps in the intelligent customer service problem matching method described above are implemented.
The embodiment of the invention has the following beneficial effects:
in the scheme, sentence vectors obtained by a reinforced self-attention model and vectors obtained by a multi-layer bidirectional Bi L STM model based on attention interaction are fully interacted, a correlation matrix between words is established, correlation scores between sentences obtained in two modes are calculated, final expression of the sentence vectors is obtained through BI L STM, self-attention mechanism analysis based on sentences and stacked bidirectional L STM based on sentence interaction are fused, sentence expression based on context is fully learned, sentence matching is achieved, compared with a keyword matching method in the prior art, the method has stronger adaptivity and flexibility, sentences which are different in expression but same in meaning can be accurately judged from the aspect of semantic understanding, sentence representation can be fully learned from more dimensions, and problem matching calculation of a customer service system is more accurate.
Drawings
FIG. 1 is a schematic flow chart of an intelligent customer service problem matching method according to an embodiment of the present invention;
FIG. 2 is a block diagram of an intelligent customer service problem matching apparatus according to an embodiment of the present invention;
FIG. 3 is a diagram illustrating an embodiment of updating a candidate question-answer library;
FIG. 4 is a schematic diagram of a problem matching algorithm according to an embodiment of the present invention;
FIG. 5 is a block diagram of an enhanced self-attention model.
Detailed Description
In order to make the technical problems, technical solutions and advantages to be solved by the embodiments of the present invention clearer, the following detailed description will be given with reference to the accompanying drawings and specific embodiments.
The names and abbreviations of the terms related to the present invention may be changed correspondingly, and the technical solution of the present invention is still applicable when the abbreviations are changed.
The embodiment of the invention provides an intelligent customer service problem matching method and device, which can improve the accuracy of a search result of an intelligent customer service system.
The embodiment of the invention provides an intelligent customer service problem matching method, as shown in fig. 1, comprising the following steps:
step 101: inputting initial word vectors of sentences 1 and sentences 2 in the candidate question-answering library into a pre-trained reinforced self-attention model to respectively obtain sentence vectors 1 and sentence vectors 2;
102, inputting sentences 1 and sentences 2 in the candidate question-answering library into a multilayer bidirectional Bi L STM model based on attention interaction to respectively obtain sentence vectors 3 and sentence vectors 4;
103, obtaining association scores between words of the sentences 1 and 2 and association scores between the words of the sentences 2 and 1 according to the sentence vectors 1, 2, 3 and 4, inputting the association scores into a BI L STM model to obtain final sentence vectors 1 and 2 of the sentences 1 and 2 respectively;
step 104: obtaining the similarity between sentence 1 and sentence 2 from the final sentence vector 1 and the final sentence vector 2 and the sentence vector 1 and the sentence vector 2;
step 105: and determining answers matched with the user questions according to the similarity between the sentences.
The method comprises the steps of taking a sentence 1 as H, taking a sentence 2 as P.H and P, obtaining vectors A (H) and A (P) through a trained self-attention-enhancing model on the one hand, obtaining B (H) and B (P) through a multi-layer bidirectional Bi L STM model on the other hand, calculating association scores align (H) between words of the sentence 1 and the sentence 2 and association scores align (P) between the words of the sentence 2 and the sentence 2 according to A (H), A (P), B (H) and B (P), and inputting the association scores align (H) and align (P) into a BI L STM to obtain final vectors F (H) and F (P) of the sentence 1.
In the embodiment, sentence vectors obtained by the strengthened self-attention model and vectors obtained by the multi-layer bidirectional Bi L STM model based on attention interaction are fully interacted, an association matrix between words is established, association scores between sentences obtained in two modes are calculated, final expression of the sentence vectors is obtained through BI L STM, self-attention mechanism analysis based on sentences and stacked bidirectional L STM based on sentence interaction are fused, sentence expression based on context is fully learned, sentence matching is achieved, compared with a keyword matching method in the prior art, the method has stronger adaptivity and flexibility, sentences which are different in expression but same in meaning can be accurately judged from the aspect of semantic understanding, sentence representation can be fully learned from more dimensions, and problem matching calculation of a customer service system is more accurate.
Further, the method further comprises:
obtaining user feedback;
and storing the questions and answers fed back by the user as satisfactory questions and answers in the candidate question and answer library.
Further, the method further comprises:
dividing the candidate question-answer library into at least one standard answer pool, wherein each standard answer pool contains a standard question corresponding to an answer and a plurality of user questions;
and taking each standard answer pool as a category, and performing multi-classification learning through a self-attention model based on reinforcement learning.
Further, the multi-classification learning through the self-attention model based on reinforcement learning includes:
dividing the user question sentences in the candidate question-answering library into two parts, respectively carrying out word embedding mapping on the first half section and the second half section of the sentence, and respectively generating a sentence vector A and a sentence vector B through the enhanced self-attention model;
and connecting sentence vectors A, B, A + B and A ⊙ B, and optimizing a multi-class cross entropy objective function by using a standard answer pool to which the user question sentence belongs as a label through a full connection layer and softmax.
Further, outputting a sentence vector using the augmented self-attention model comprises:
inputting a sentence vector x ═ x subjected to embedding mapping1,...,xn]；
Respectively obtaining two vectors with the length of n and composed of 1 and 0 through preset probability distribution sampling
Generating a mask matrix:
the score f (x) between the elements of the vector xi,xj) Formed matrix and mask matrix Mij rssBy element addition, a self-attention score f is obtainedrss(xi,xj) For each xjThe softmax operation is carried out and,
to obtain
Deriving a context vector
Combining s and x through a fusion gate to obtain a final sentence vector u;
wherein F is sigmoid (W)(f)[x；s]+b(f)),u＝F⊙x+(1-F)⊙s。
Furthermore, the multilayer bidirectional Bi L STM model consists of 4 layers of Bi L STMs combined with attention attentions, the input of each layer of Bi L STM is formed by connecting the output of the previous layer of hidden layer, the output of the previous layer of attention model and the input of the previous layer, and the hidden layer input of each layer consists of the current input and the hidden layer output of the current layer at the previous moment.
Further, the step of inputting sentences 1 and sentences 2 in the candidate question-answering library into the attention interaction-based multi-layer bidirectional Bi L STM model to obtain sentence vectors 3 and sentence vectors 4 respectively comprises:
inputting word vectors, part-of-speech POS vectors, named entity recognition NER vectors and accurate matching EM vectors of the sentence 1 and the sentence 2 in the candidate question-answering library into the multilayer bidirectional Bi L STM model to obtain a correlation matrix of each word in the sentence 1 and the sentence 2;
and inputting the obtained incidence matrix into a self-coding layer, performing preset feature dimension reduction, and obtaining a sentence vector 3 and a sentence vector 4 through a max-posing layer.
Further, the obtaining of association scores between words of the sentence 1 and the sentence 2 and association scores between the words of the sentence 2 and the sentences 1 according to the sentence vector 1, the sentence vector 2, the sentence vector 3 and the sentence vector 4, and the inputting of the association scores into a BI L STM model to obtain a final sentence vector 1 and a final sentence vector 2 of the sentence 1 and the sentence 2 respectively comprises:
connecting sentence vector 1, sentence vector 2, sentence vector 3 and sentence vector 4 to obtain a similarity matrix;
calculating the association score generated by each word in sentence 1 to each word in sentence 2 according to the similarity matrix;
and inputting the association score into a layer of Bi L STM model to obtain a final sentence vector 1 and a final sentence vector 2 of the sentence 1 and the sentence 2.
Further, the obtaining the similarity between sentence 1 and sentence 2 by using the final sentence vector 1 and final sentence vector 2 and the sentence vector 1 and sentence vector 2 includes:
connecting the final sentence vector 1 and the final sentence vector 2 with the sentence vector 1 and the sentence vector 2 to form a vector [ final sentence vector 1, final sentence vector 2, final sentence vector 1 ⊙, final sentence vector 2, sentence vector 1 ⊙ sentence vector 2 ];
the similarity between sentence 1 and sentence 2 is obtained through the full-connected layer and softmax.
An embodiment of the present invention further provides an intelligent customer service problem matching device, as shown in fig. 2, including:
the first processing module 21 is configured to input initial word vectors of sentences 1 and sentences 2 in the candidate question-answering library into a pre-trained reinforced self-attention model, so as to obtain sentence vectors 1 and sentence vectors 2, respectively;
the second processing module 22 is configured to input sentences 1 and sentences 2 in the candidate question-answering library into a multi-layer bidirectional Bi L STM model based on attention interaction to obtain sentence vectors 3 and sentence vectors 4, respectively;
the third processing module 23 is configured to obtain association scores between words of the sentence 1 and the sentence 2 and association scores between the sentence 2 and the words of the sentence 1 according to the sentence vector 1, the sentence vector 2, the sentence vector 3 and the sentence vector 4, and input the association scores into a BI L STM model to obtain a final sentence vector 1 and a final sentence vector 2 of the sentence 1 and the sentence 2, respectively;
the similarity calculation module 24 is configured to obtain similarities between the final sentence vector 1 and the final sentence vector 2 and the sentence vectors 1 and 2;
and a matching module 25, configured to determine an answer matched with the user question according to the similarity between the sentences.
In the embodiment, sentence vectors obtained by the strengthened self-attention model and vectors obtained by the multi-layer bidirectional Bi L STM model based on attention interaction are fully interacted, an association matrix between words is established, association scores between sentences obtained in two modes are calculated, final expression of the sentence vectors is obtained through BI L STM, self-attention mechanism analysis based on sentences and stacked bidirectional L STM based on sentence interaction are fused, sentence expression based on context is fully learned, sentence matching is achieved, compared with a keyword matching method in the prior art, the method has stronger adaptivity and flexibility, sentences which are different in expression but same in meaning can be accurately judged from the aspect of semantic understanding, sentence representation can be fully learned from more dimensions, and problem matching calculation of a customer service system is more accurate.
Further, the apparatus further comprises:
the storage module is used for acquiring user feedback; and storing the questions and answers fed back by the user as satisfactory questions and answers in the candidate question and answer library.
Further, the apparatus further comprises:
the training module is used for dividing the candidate question-answer library into at least one standard answer pool, and each standard answer pool contains standard questions corresponding to answers and a plurality of user questions; and taking each standard answer pool as a category, and performing multi-classification learning through a self-attention model based on reinforcement learning.
Further, the training module is specifically configured to:
dividing the user question sentences in the candidate question-answering library into two parts, respectively carrying out word embedding mapping on the first half section and the second half section of the sentence, and respectively generating a sentence vector A and a sentence vector B through the enhanced self-attention model;
and connecting sentence vectors A, B, A + B and A ⊙ B, and optimizing a multi-class cross entropy objective function by using a standard answer pool to which the user question sentence belongs as a label through a full connection layer and softmax.
Further, the first processing module is specifically configured to output a sentence vector using the augmented self-attention model by:
inputting a sentence vector x ═ x subjected to embedding mapping1,...,xn]；
Respectively obtaining two vectors with the length of n and composed of 1 and 0 through preset probability distribution sampling
Generating a mask matrix:
the score f (x) between the elements of the vector xi,xj) Formed matrix and mask matrix Mij rssBy element addition, a self-attention score f is obtainedrss(xi,xj) For each xjThe softmax operation is carried out and,
to obtain
Deriving a context vector
Combining s and x through a fusion gate to obtain a final sentence vector u;
wherein F is sigmoid (W)(f)[x；s]+b(f)),u＝F⊙x+(1-F)⊙s。
Furthermore, the multilayer bidirectional Bi L STM model consists of 4 layers of Bi L STMs combined with attention attentions, the input of each layer of Bi L STM is formed by connecting the output of the previous layer of hidden layer, the output of the previous layer of attention model and the input of the previous layer, and the hidden layer input of each layer consists of the current input and the hidden layer output of the current layer at the previous moment.
The second processing module is specifically configured to input word vectors, part-of-speech tagging POS vectors, named entity recognition NER vectors, and exact matching EM vectors of sentences 1 and 2 in the candidate question-answering library into the multi-layer bidirectional Bi L STM model to obtain an incidence matrix of each word in the sentence 1 and the sentence 2, input the obtained incidence matrix into a self-coding layer, perform preset feature dimension reduction, and obtain a sentence vector 3 and a sentence vector 4 through a max-posing layer.
The third processing module is specifically configured to link the sentence vector 1, the sentence vector 2, the sentence vector 3, and the sentence vector 4 to obtain a similarity matrix, calculate a relevance score generated by each word in the sentence 1 for each word in the sentence 2 according to the similarity matrix, and input the relevance score into a Bi L STM model to obtain a final sentence vector 1 and a final sentence vector 2 of the sentence 1 and the sentence 2.
Further, the similarity calculation module is specifically configured to connect the final sentence vector 1 and the final sentence vector 2 with the sentence vector 1 and the sentence vector 2 to form vectors [ the final sentence vector 1; the final sentence vector 2; the final sentence vector 1 ⊙; the final sentence vector 2; the sentence vector 1 ⊙; and the sentence vector 2], and obtain the similarity between the sentences 1 and 2 through a full connection layer and softmax.
The technical scheme of the invention is further described by combining the drawings and specific embodiments:
the embodiment provides a problem matching method and device based on an enhanced self-attention model. In this embodiment, a mechanism for expanding the candidate question-answer library according to the user feedback is added to the customer service system, and as shown in fig. 3, a satisfactory question-answer pair fed back by the user can be added to the candidate question-answer library in real time, so that the number and quality of the candidate question-answer library are continuously improved, and the adaptability of the intelligent customer service system is further enhanced.
Compared with the existing intelligent customer service system, the candidate question-answer library is continuously expanded through an automatic process, so that the questions corresponding to the same standard answer are diversified, and the phenomenon that the standard questions cannot be matched due to the difference of the questions asked by the user and different expressions of the same service word is reduced. The conversation questions and the user questions in the candidate question-answer library belong to languages spontaneously organized by the user, and the mode has certain similarity, so that the two questions can share one sentence coding model, and the complexity of the algorithm is effectively reduced. The embodiment dynamically expands the candidate question-answer library, continuously supplements the question-answer pairs satisfying the user into the candidate question-answer library by recording the positive feedback of the user, enables the sentences of the candidate question-answer library to continuously trend the question way of the user, can more synchronously extract the sentence characteristics of the conversation questions and the candidate question-answer library questions when matching the questions, can also enable the training data to continuously increase and update the user questions, has good practical effectiveness, and avoids the problem that the accuracy of the system is inevitably reduced along with the expansion of the user group caused by the fixed question-answer library in the prior art.
In the existing technical scheme, the problem of expression difference caused by billions of customer visits is far from being solved through the keyword matching problem. The problem matching algorithm based on deep learning can effectively utilize the advantage of large data volume, and the algorithm trend is to encode sentences by optimizing a target function in a specific task, obtain vector expression of the sentences and perform matching. In the prior art, only a model which simply performs sentence vectorization calculation and matching through a recurrent neural network or a convolutional neural network exists, and the algorithm for sentence coding through the neural network has low accuracy due to the fixed defect of a single model.
In order to solve the problem, the embodiment is based on the idea of migration learning and model fusion, a reinforced self-attention model obtained by training a classification task is migrated into a similarity measurement task, and a more comprehensive sentence representation vector is generated through information interaction between vectors.
The algorithm flow of the present embodiment is shown in fig. 4, wherein the algorithm framework is divided into two parts: the method comprises the steps of classification task-based reinforced self-attention model learning and similarity task-based multi-model fusion learning.
Wherein, the reinforced self-attention model based on the classification task is described as follows:
(1) preparing data: dividing the candidate question-answer library into a plurality of standard answer pools, wherein each standard answer pool comprises standard questions corresponding to answers and a large number of user questions. And taking each standard answer pool as a category, and performing multi-classification learning through a self-attention model based on reinforcement learning.
(2) Model description:
the framework of the self-attention-enhancing model is shown in fig. 5, where an embeding mapped sentence vector x ═ x is input1,...,xn]Outputting a context-dependent sentence vector expression u ═ u1,...,un]。
Firstly, two vectors with the length of n and composed of 1 and 0 are respectively obtained through certain probability distribution sampling
Both generate a mask matrix
The score f (x) between the elements of the vector xi,xj) Formed matrix and mask matrix Mij rssBy element addition, a self-attention score f is obtainedrss(xi,xj) For each xjThe softmax operation is carried out and,
to obtain
Thereby obtainingContext vector
And combining s and x through a fusion gate to obtain a final sentence vector u.
F＝sigmoid(W(f)[x；s]+b(f)),u＝F⊙x+(1-F)⊙s
(3) Data input: dividing a user question sentence into two parts, performing wordemtwodding mapping on the first half part and the second half part of the sentence respectively, and generating sentence vectors A and B through an enhanced self-attention model respectively.
(4) And connecting the vectors A, B, A + B and A ⊙ B, and optimizing the multi-class cross entropy objective function by using the standard answer pool to which the sentence belongs as label through the full connection layer and softmax.
(5) And storing the reinforced self-attention model and the model parameters obtained by the multi-classification model learning. For further migratory learning.
The multi-model fusion learning model based on the similarity task is described as follows:
(1) the input of the neural network is formed by connecting five parts: word vectors, part-of-speech tagged (POS) vectors, Named Entity Recognition (NER) vectors, Exact Match (EM) vectors.
(2) The Word vector and the Word vector are obtained by mapping sentences and words in a trained Word2vec vector table after being divided according to words, wherein the Word vector can solve the problem of unknown words to a certain extent.
(3) The part-of-speech tagging vector and the named entity recognition vector are obtained by searching vectors corresponding to vocabularies through a trained vector table in the part-of-speech tagging task and the named entity recognition task, and the two vectors represent sentences from the perspective of grammar and sentence structures.
(4) An exact match characterizes whether the current word is in another sentence by a 1 or 0.
(5) Input freq(x)＝[word2vec(x)；char(x)；POS(x)；NER(x)；EM(x)]Wherein word2vec (x) is a word vector, char (x) is a word vector, pos (x) is a part-of-speech tagging vector, ner (x) is a named entity identification vector, and em (x) is an exact match vector.
Pi＝BiLSTM(frep(Pi),Pi-1),i＝1,...,n，uj＝BiLSTM(frep(Hj),uj-1),j＝1,...,n
(6) the Attenttion layer calculates the association score of each word in P with sentence H using the following formula
(7) The attention interaction-based multi-layer bidirectional Bi L STM model consists of 4 layers of Bi L STMs combined with attentions, the input of each layer of Bi L STM is formed by connecting the output of the previous layer of hidden layer, the output of the previous layer of attention model and the input of the previous layer, and the hidden layer input of each layer consists of the current input and the hidden layer output of the current layer at the previous moment.
Wherein,
for the output of the hidden layer of the layer l at the time t,
for the output of the hidden layer at the time of t-1, Hl() For the function of the hidden layer of the layer l,
for the output of the layer 1 attention model at time t,
for the layer input at time t and l,
for the l-1 level input at time t.
(8) Inputting a matrix obtained by combining Bi L STM of authentication by 4 layers into a self-coding layer to perform certain characteristic dimension reduction, and obtaining a sentence vector P through a max-posing layerv,Hv。
(9) Inputting the initial word vectors of the sentence 1 and the sentence 2 into a pre-trained reinforced self-attention model to obtain a sentence vector 1 and a sentence vector 2 which are respectively represented as rp,rh。
(10) Will Pv,Hv,rp,rhThe connection results in a similarity matrix A which is,
(11) Calculate the ith word in sentence 1 using the following formula
The entries generated for each word in H,
(12) f is a layer of activation function Re L u, will
And
inputting into a Bi L STM to obtain
As a vector that ultimately characterizes the sentence.
(13) Vector p that will characterize a sentenceM,uMAnd rp,rhConnected to form a vector [ pM；uM；pM⊙uM；rp⊙rh]And finally, obtaining the similarity prediction through one-layer full connection and softmax.
In the embodiment, a classification task is used for pre-training a reinforced self-attention model, the trained model is transferred to a similarity prediction task for sentence coding, and different representations of semantics and grammar angles such as word vectors, part of speech tagging, named entities, precise matching and the like are connected to be used as model input in combination with a Bi L STM based on an attention model, so that the input sentence characteristics are enriched, the network can learn sentences more comprehensively and more specifically, a problem matching method of multi-model fusion is realized, and a problem matching method with high accuracy is realized by integrating the self semantics and grammar characteristics of sentences, the correlation between sentences and the like.
The method is characterized in that a self-attention model based on reinforcement learning is combined with an attention model connected in a bidirectional L STM, the advantages of less parameters of the attention model and strong learning capacity are fully utilized, the attention mechanism of the self-attention model of the concerned sentence and the concerned sentence are comprehensively combined, partial data in an attention matrix is adopted, and a soft attention algorithm with strong fault tolerance and generalization capacity is realized.
The embodiment of the invention also provides an intelligent customer service problem matching device, which comprises: a memory, a processor and a computer program stored on the memory and executable on the processor, the computer program when executed by the processor implementing the steps in the intelligent customer service problem matching method as described above.
An embodiment of the present invention further provides a computer-readable storage medium, where a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the steps in the intelligent customer service problem matching method described above are implemented.
For a hardware implementation, the Processing units may be implemented within one or more Application Specific Integrated Circuits (ASICs), Digital Signal Processors (DSPs), Digital Signal Processing Devices (DSPDs), Programmable logic devices (P L D), Field-Programmable Gate arrays (FPGAs), general purpose processors, controllers, microcontrollers, microprocessors, other electronic units configured to perform the functions described herein, or a combination thereof.
For a software implementation, the techniques described herein may be implemented with modules (e.g., procedures, functions, and so on) that perform the functions described herein. The software codes may be stored in a memory and executed by a processor. The memory may be implemented within the processor or external to the processor.
The embodiments in the present specification are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other.
As will be appreciated by one skilled in the art, embodiments of the present invention may be provided as a method, apparatus, or computer program product. Accordingly, embodiments of the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, embodiments of the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
Embodiments of the present invention are described with reference to flowchart illustrations and/or block diagrams of methods, user equipment (systems), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
While preferred embodiments of the present invention have been described, additional variations and modifications of these embodiments may occur to those skilled in the art once they learn of the basic inventive concepts. Therefore, it is intended that the appended claims be interpreted as including preferred embodiments and all such alterations and modifications as fall within the scope of the embodiments of the invention.
It is further noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or user equipment that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or user equipment. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other like elements in a process, method, article, or user equipment that comprises the element.
While the preferred embodiments of the present invention have been described, it will be understood by those skilled in the art that various changes and modifications may be made without departing from the spirit and scope of the invention as defined in the following claims.
Claims (20)
1. An intelligent customer service problem matching method is characterized by comprising the following steps:
inputting initial word vectors of sentences 1 and sentences 2 in the candidate question-answering library into a pre-trained reinforced self-attention model to respectively obtain sentence vectors 1 and sentence vectors 2;
inputting sentences 1 and sentences 2 in the candidate question-answering library into a multilayer bidirectional Bi L STM model based on attention interaction to respectively obtain sentence vectors 3 and sentence vectors 4;
obtaining association scores between words of the sentences 1 and 2 and association scores between the words of the sentences 2 and 1 according to the sentence vectors 1, 2, 3 and 4, inputting the association scores into a BI L STM model to obtain final sentence vectors 1 and 2 of the sentences 1 and 2 respectively;
obtaining the similarity between sentence 1 and sentence 2 from the final sentence vector 1 and the final sentence vector 2 and the sentence vector 1 and the sentence vector 2;
and determining answers matched with the user questions according to the similarity between the sentences.
2. The intelligent customer service problem matching method according to claim 1, further comprising:
obtaining user feedback;
and storing the questions and answers fed back by the user as satisfactory questions and answers in the candidate question and answer library.
3. The intelligent customer service problem matching method according to claim 1, further comprising:
dividing the candidate question-answer library into at least one standard answer pool, wherein each standard answer pool contains a standard question corresponding to an answer and a plurality of user questions;
and taking each standard answer pool as a category, and performing multi-classification learning through a self-attention model based on reinforcement learning.
4. The intelligent customer service problem matching method according to claim 3, wherein the multi-classification learning through the reinforcement learning-based self-attention model comprises:
dividing the user question sentences in the candidate question-answering library into two parts, performing wordemtwodding mapping on the first half section and the second half section of the sentence respectively, and generating sentence vectors A and sentence vectors B through the enhanced self-attention model respectively;
and connecting sentence vectors A, B, A + B and A ⊙ B, and optimizing a multi-class cross entropy objective function by using a standard answer pool to which the user question sentence belongs as a label through a full connection layer and softmax.
5. The intelligent customer service question matching method of claim 1 wherein outputting a sentence vector using the augmented self-attention model comprises:
inputting a sentence vector x ═ x subjected to embedding mapping1,...,xn]；
Respectively obtaining two vectors with the length of n and composed of 1 and 0 through preset probability distribution sampling
Generating a mask matrix:
the score f (x) between the elements of the vector xi,xj) Formed matrix and mask matrix Mij rssBy element addition, a self-attention score f is obtainedrss(xi,xj) For each xjThe softmax operation is carried out and,
to obtain
Deriving a context vector
Combining s and x through a fusion gate to obtain a final sentence vector u;
wherein F is sigmoid (W)(f)[x；s]+b(f)),u＝F⊙x+(1-F)⊙s。
6. The intelligent customer service problem matching method according to claim 1, wherein the multi-layer bidirectional Bi L STM model consists of 4 layers of Bi L STMs combined with attention attentions, the input of each layer of Bi L STM is formed by connecting the output of the previous layer of hidden layer, the output of the previous layer of attention model and the input of the previous layer, and the hidden layer input of each layer consists of the current input and the hidden layer output of the current layer at the previous moment.
7. The intelligent customer service question matching method according to claim 6, wherein the step of inputting sentences 1 and sentences 2 in the candidate question-and-answer library into the attention interaction based multi-layer bidirectional Bi L STM model to obtain sentence vectors 3 and sentence vectors 4 comprises the steps of:
inputting word vectors, part-of-speech POS vectors, named entity recognition NER vectors and accurate matching EM vectors of the sentence 1 and the sentence 2 in the candidate question-answering library into the multilayer bidirectional Bi L STM model to obtain a correlation matrix of each word in the sentence 1 and the sentence 2;
and inputting the obtained incidence matrix into a self-coding layer, performing preset feature dimension reduction, and obtaining a sentence vector 3 and a sentence vector 4 through a max-posing layer.
8. The intelligent customer service problem matching method according to claim 1, wherein the obtaining of association scores between words of sentence 1 and sentence 2 and association scores between words of sentence 2 and sentence 1 from sentence vector 1, sentence vector 2, sentence vector 3 and sentence vector 4 and the inputting of the association scores into the BI L STM model to obtain final sentence vector 1 and final sentence vector 2 of sentence 1 and sentence 2 respectively comprises:
connecting sentence vector 1, sentence vector 2, sentence vector 3 and sentence vector 4 to obtain a similarity matrix;
calculating the association score generated by each word in sentence 1 to each word in sentence 2 according to the similarity matrix;
and inputting the association score into a layer of Bi L STM model to obtain a final sentence vector 1 and a final sentence vector 2 of the sentence 1 and the sentence 2.
9. The intelligent customer service question matching method according to claim 1, wherein the obtaining of the similarity between sentence vector 1 and sentence vector 2 and sentence vector 1 and sentence vector 2 comprises:
connecting the final sentence vector 1 and the final sentence vector 2 with the sentence vector 1 and the sentence vector 2 to form a vector [ final sentence vector 1, final sentence vector 2, final sentence vector 1 ⊙, final sentence vector 2, sentence vector 1 ⊙ sentence vector 2 ];
the similarity between sentence 1 and sentence 2 is obtained through the full-connected layer and softmax.
10. An intelligent customer service problem matching device, comprising:
the first processing module is used for inputting initial word vectors of sentences 1 and sentences 2 in the candidate question-answering library into a pre-trained reinforced self-attention model to respectively obtain sentence vectors 1 and sentence vectors 2;
the second processing module is used for inputting sentences 1 and sentences 2 in the candidate question-answering library into a multilayer bidirectional Bi L STM model based on attention interaction to respectively obtain sentence vectors 3 and sentence vectors 4;
the third processing module is used for obtaining association scores between words of the sentence 1 and the sentence 2 and association scores between the words of the sentence 2 and the sentences 1 according to the sentence vector 1, the sentence vector 2, the sentence vector 3 and the sentence vector 4, inputting the association scores into a BI L STM model to respectively obtain a final sentence vector 1 and a final sentence vector 2 of the sentence 1 and the sentence 2;
the similarity calculation module is used for obtaining the similarity between the sentence vector 1 and the sentence vector 2 from the final sentence vector 1 and the final sentence vector 2 and the sentence vector 1 and the sentence vector 2;
and the matching module is used for determining answers matched with the user questions according to the similarity between the sentences.
11. The intelligent customer service problem matching device according to claim 10, further comprising:
the storage module is used for acquiring user feedback; and storing the questions and answers fed back by the user as satisfactory questions and answers in the candidate question and answer library.
12. The intelligent customer service problem matching device according to claim 10, further comprising:
the training module is used for dividing the candidate question-answer library into at least one standard answer pool, and each standard answer pool contains standard questions corresponding to answers and a plurality of user questions; and taking each standard answer pool as a category, and performing multi-classification learning through a self-attention model based on reinforcement learning.
13. The intelligent customer service problem matching device of claim 12, wherein the training module is specifically configured to:
dividing the user question sentences in the candidate question-answering library into two parts, performing wordemtwodding mapping on the first half section and the second half section of the sentence respectively, and generating sentence vectors A and sentence vectors B through the enhanced self-attention model respectively;
and connecting sentence vectors A, B, A + B and A ⊙ B, and optimizing a multi-class cross entropy objective function by using a standard answer pool to which the user question sentence belongs as a label through a full connection layer and softmax.
14. The intelligent customer service problem matching device according to claim 10, wherein the first processing module is specifically configured to output sentence vectors using the augmented self-attention model by:
inputting a sentence vector x ═ x subjected to embedding mapping1,...,xn]；
Respectively obtaining two vectors with the length of n and composed of 1 and 0 through preset probability distribution sampling
Generating a mask matrix:
the score f (x) between the elements of the vector xi,xj) Formed matrix and mask matrix Mij rssBy element addition, a self-attention score f is obtainedrss(xi,xj) For each xjThe softmax operation is carried out and,
to obtain
Deriving a context vector
Combining s and x through a fusion gate to obtain a final sentence vector u;
wherein F is sigmoid (W)(f)[x；s]+b(f)),u＝F⊙x+(1-F)⊙s。
15. The intelligent customer service problem matching device according to claim 10, wherein the multi-layer bidirectional Bi L STM model is composed of 4 layers of Bi L STMs combined with attention attentions, the input of each layer of Bi L STM is formed by connecting the output of the previous layer of hidden layer, the output of the previous layer of attention model and the input of the previous layer, and the hidden layer input of each layer is composed of the current input and the hidden layer output of the current layer at the previous moment.
16. The intelligent customer service problem matching device of claim 15,
the second processing module is specifically used for inputting word vectors, part-of-speech tagging POS vectors, named entity recognition NER vectors and accurate matching EM vectors of the sentences 1 and 2 in the candidate question-answering library into the multilayer bidirectional Bi L STM model to obtain a correlation matrix of each word and each sentence 2 in the sentences 1, inputting the obtained correlation matrix into a coding layer, performing preset feature dimension reduction, and obtaining sentence vectors 3 and sentence vectors 4 through a max-posing layer.
17. The intelligent customer service problem matching device of claim 10,
the third processing module is specifically used for connecting the sentence vector 1, the sentence vector 2, the sentence vector 3 and the sentence vector 4 to obtain a similarity matrix, calculating a relevance score generated by each word in the sentence 1 to each word in the sentence 2 according to the similarity matrix, and inputting the relevance score into a layer of Bi L STM model to obtain a final sentence vector 1 and a final sentence vector 2 of the sentence 1 and the sentence 2.
18. The intelligent customer service problem matching device of claim 10,
the similarity calculation module is specifically used for connecting the final sentence vector 1 and the final sentence vector 2 with the sentence vector 1 and the sentence vector 2 to form a vector [ the final sentence vector 1, the final sentence vector 2, the final sentence vector 1 ⊙, the final sentence vector 2, the sentence vector 1 ⊙, the sentence vector 2], and obtaining the similarity between the sentences 1 and 2 through a full connection layer and softmax.
19. An intelligent customer service problem matching device, comprising: memory, processor and computer program stored on the memory and executable on the processor, which when executed by the processor implements the steps in the intelligent customer service problem matching method according to any one of claims 1 to 9.
20. A computer-readable storage medium, having stored thereon a computer program which, when executed by a processor, carries out the steps in the intelligent customer service problem matching method of any one of claims 1 to 9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910000948.XA CN111400461B (en) | 2019-01-02 | 2019-01-02 | Intelligent customer service problem matching method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910000948.XA CN111400461B (en) | 2019-01-02 | 2019-01-02 | Intelligent customer service problem matching method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111400461A true CN111400461A (en) | 2020-07-10 |
| CN111400461B CN111400461B (en) | 2023-03-31 |
Family
ID=71433929
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910000948.XA Active CN111400461B (en) | 2019-01-02 | 2019-01-02 | Intelligent customer service problem matching method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111400461B (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111858893A (en) * | 2020-07-27 | 2020-10-30 | 平安科技（深圳）有限公司 | Sentence pair matching method and device, computer equipment and storage medium |
| CN112052319A (en) * | 2020-09-01 | 2020-12-08 | 杭州师范大学 | Intelligent customer service method and system based on multi-feature fusion |
| CN112182231A (en) * | 2020-12-01 | 2021-01-05 | 佰聆数据股份有限公司 | Text processing method, system and storage medium based on sentence vector pre-training model |
| CN112215627A (en) * | 2020-12-09 | 2021-01-12 | 强链(江苏)科创发展有限公司 | Customer information data processing system |
| CN112861071A (en) * | 2021-02-05 | 2021-05-28 | 哈尔滨工程大学 | High-speed rail traction system anomaly detection method based on deep self-coding |
| CN113239678A (en) * | 2021-04-02 | 2021-08-10 | 南京邮电大学 | Multi-angle attention feature matching method and system for answer selection |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180181673A1 (en) * | 2016-12-28 | 2018-06-28 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Answer searching method and device based on deep question and answer |
| CN108764194A (en) * | 2018-06-04 | 2018-11-06 | 科大讯飞股份有限公司 | A kind of text method of calibration, device, equipment and readable storage medium storing program for executing |
| CN108829719A (en) * | 2018-05-07 | 2018-11-16 | 中国科学院合肥物质科学研究院 | The non-true class quiz answers selection method of one kind and system |
| US20180349359A1 (en) * | 2017-05-19 | 2018-12-06 | salesforce.com,inc. | Natural language processing using a neural network |
-
2019
- 2019-01-02 CN CN201910000948.XA patent/CN111400461B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180181673A1 (en) * | 2016-12-28 | 2018-06-28 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Answer searching method and device based on deep question and answer |
| US20180349359A1 (en) * | 2017-05-19 | 2018-12-06 | salesforce.com,inc. | Natural language processing using a neural network |
| CN108829719A (en) * | 2018-05-07 | 2018-11-16 | 中国科学院合肥物质科学研究院 | The non-true class quiz answers selection method of one kind and system |
| CN108764194A (en) * | 2018-06-04 | 2018-11-06 | 科大讯飞股份有限公司 | A kind of text method of calibration, device, equipment and readable storage medium storing program for executing |
Non-Patent Citations (1)
| Title |
|---|
| 张小艳: "中文主观题自动批改中相似句子检索算法", 《南京师范大学学报(工程技术版)》 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111858893A (en) * | 2020-07-27 | 2020-10-30 | 平安科技（深圳）有限公司 | Sentence pair matching method and device, computer equipment and storage medium |
| CN111858893B (en) * | 2020-07-27 | 2022-06-03 | 平安科技（深圳）有限公司 | Sentence pair matching method and device, computer equipment and storage medium |
| CN112052319A (en) * | 2020-09-01 | 2020-12-08 | 杭州师范大学 | Intelligent customer service method and system based on multi-feature fusion |
| CN112052319B (en) * | 2020-09-01 | 2022-05-17 | 杭州师范大学 | Intelligent customer service method and system based on multi-feature fusion |
| CN112182231A (en) * | 2020-12-01 | 2021-01-05 | 佰聆数据股份有限公司 | Text processing method, system and storage medium based on sentence vector pre-training model |
| CN112182231B (en) * | 2020-12-01 | 2021-03-09 | 佰聆数据股份有限公司 | Text processing method, system and storage medium based on sentence vector pre-training model |
| CN112215627A (en) * | 2020-12-09 | 2021-01-12 | 强链(江苏)科创发展有限公司 | Customer information data processing system |
| CN112861071A (en) * | 2021-02-05 | 2021-05-28 | 哈尔滨工程大学 | High-speed rail traction system anomaly detection method based on deep self-coding |
| CN113239678A (en) * | 2021-04-02 | 2021-08-10 | 南京邮电大学 | Multi-angle attention feature matching method and system for answer selection |
| CN113239678B (en) * | 2021-04-02 | 2023-06-20 | 南京邮电大学 | Multi-angle attention feature matching method and system for answer selection |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111400461B (en) | 2023-03-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111400461B (en) | Intelligent customer service problem matching method and device | |
| CN110175227B (en) | Dialogue auxiliary system based on team learning and hierarchical reasoning | |
| US11862145B2 (en) | Deep hierarchical fusion for machine intelligence applications | |
| CN113420807A (en) | Multi-mode fusion emotion recognition system and method based on multi-task learning and attention mechanism and experimental evaluation method | |
| Wen et al. | Dynamic interactive multiview memory network for emotion recognition in conversation | |
| CN111460132B (en) | Generation type conference abstract method based on graph convolution neural network | |
| CN114596844B (en) | Training method of acoustic model, voice recognition method and related equipment | |
| CN112417894A (en) | Conversation intention identification method and system based on multi-task learning | |
| CN112559706B (en) | Training method of dialogue generating model, dialogue method, device and storage medium | |
| Cabada et al. | Mining of educational opinions with deep learning | |
| Qian et al. | A prompt-aware neural network approach to content-based scoring of non-native spontaneous speech | |
| Sharath et al. | Question answering over knowledge base using language model embeddings | |
| Cont | Modeling musical anticipation: From the time of music to the music of time | |
| Chandiok et al. | CIT: Integrated cognitive computing and cognitive agent technologies based cognitive architecture for human-like functionality in artificial systems | |
| CN116821294A (en) | Question-answer reasoning method and device based on implicit knowledge ruminant | |
| Aina et al. | What do entity-centric models learn? insights from entity linking in multi-party dialogue | |
| CN117037789B (en) | Customer service voice recognition method and device, computer equipment and storage medium | |
| CN106503066A (en) | Process Search Results method and apparatus based on artificial intelligence | |
| CN115171870A (en) | Diagnosis guiding and prompting method and system based on m-BERT pre-training model | |
| Han et al. | Generative adversarial networks for open information extraction | |
| CN116737911A (en) | Deep learning-based hypertension question-answering method and system | |
| CN115169472A (en) | Music matching method and device for multimedia data and computer equipment | |
| CN114333790A (en) | Data processing method, device, equipment, storage medium and program product | |
| Su et al. | Dialog State Tracking and action selection using deep learning mechanism for interview coaching | |
| Kreyssig | Deep learning for user simulation in a dialogue system |
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 |