CN109948163A - The natural language semantic matching method that sequence dynamic is read - Google Patents
The natural language semantic matching method that sequence dynamic is read Download PDFInfo
- Publication number
- CN109948163A CN109948163A CN201910228242.9A CN201910228242A CN109948163A CN 109948163 A CN109948163 A CN 109948163A CN 201910228242 A CN201910228242 A CN 201910228242A CN 109948163 A CN109948163 A CN 109948163A
- Authority
- CN
- China
- Prior art keywords
- natural language
- vector
- word
- language sentences
- sentence
- 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
Abstract
The invention discloses a kind of natural language semantic matching methods that sequence dynamic is read, comprising: carries out semantic modeling to each word of natural language sentences centering, obtaining the corresponding word level semantics of each natural language sentences indicates vector;Vector is indicated according to each word level semantics, and corresponding sentence semantics expression vector is obtained by stacking neural network and considers complementary hidden layer expression vector between word;Vector is indicated using sentence semantics and considers the dynamic understanding that complementary hidden layer expression vector between word carries out sentence semantics, and the dynamic for obtaining corresponding sentence, which understands, indicates vector;The sentence semantics of natural language sentences pair are indicated that the dynamic of vector and sentence understands expression vector, are respectively integrated, and realize that the semantic relation of natural language sentences pair is classified according to integrated results.This method can be read by the dynamic of distich subsequence and realize the subsemantic accurate understanding of distich and expression, and then realize the accurate judgement to natural language semantic matches.
Description
Technical field
The present invention relates to oneself of deep learning and natural language understanding technology field more particularly to a kind of dynamic reading of sequence
Right language semantic matching process.
Background technique
Natural language sentences semantic matches (Sentence Semantic Matching) are natural language processing fields
The research contents of one basis but key, the main problem solved is the semantic relation judged between two sentences.Such as
In natural language inference (Natural Language Inference, NLI), sentence semantics matching is mainly used for judging hypothetical sentence
Semanteme whether can be inferred from premise sentence come.In repeating identification (Paraphrase Identification, PI), sentence
Sub- semantic matches are mainly used for judging whether two sentences are same semantic in expression.Therefore the task matter of utmost importance to be solved
It is the semantic expressiveness that therefore task matter of utmost importance to be solved is natural language sentences.The semantic table of natural language sentences
Show it is natural language processing even one basis of artificial intelligence field but extremely important research contents, either basic information
Retrieval, semantics extraction, or complicated question answering system, conversational system require have a standard comprehensively to the semanteme of input sentence
True expression just can guarantee that machine understands the language system of mankind's complexity in this way.Researcher has been proposed a variety of different
Semantic expressiveness learning method, wherein imitate the attention mechanism learning method of the Attention behavior of the mankind by more and more
Concern.Attention mechanism can help to select word those of important to semantic expressiveness in sentence, while can not be long by sentence
Dependence in the limitation modeling sentence of degree between word and word, provides important technical support for the semantic expressiveness of sentence.
More complicated, researcher proposes bull attention mechanism (Multi-Head attention), by considering different situations, from
Different angle models sentence semantics, thus realize to sentence semantics more comprehensively, it is more accurate to understand expression.Therefore, sharp
With attention mechanism to natural language semantic expressiveness carry out research have become natural language field exploration one it is particularly significant
Research direction.
Currently, mainly having the following contents to the research of natural language sentences semantic expressiveness using attention mechanism:
In biological cognitive science, attention mechanism can help people be absorbed in in the maximally related content of target.Therefore,
Mainly pass through a variety of different neural network structures, mould using method of the attention mechanism to natural language sentences semantic expressiveness
The attention method of bionical object, such as interior attention (inner-attention), bull attention (multi-head
Attention), mutual attention (co-attention), the modeling of the methods of direction attention (directional attention)
Word and word in sentence, semantic dependency and matching relationship between sentence then again by different neural network structures, such as are rolled up
Product neural network (Convolutional Neural Network, CNN), Recognition with Recurrent Neural Network (Recurrent Neural
Network, RNN) etc. integrate these information, obtain the final expression of sentence semantics, and apply it to different specific tasks
In the middle.
It is above-mentioned that the method for sentence semantic expressiveness is defaulted to the word sequence in entire sentence using the realization of attention mechanism
According to being handled from left to right or from the reading method turned left of the right side.And cognitive science is studies have shown that in real life, the mankind
There is very big difference using attention mechanism processing sequence.Researcher has found people's meeting when reading by eyeball tracking instrument
Ignore some words, and the mankind can select different focus according to that he is seen and that he intentionally gets and read suitable
Sequence.Further research has shown that the mankind only can be concerned about 1.5 words every time in deep reading, and the mankind once at most pay close attention to
The target different to 7.These all demonstrate human attention mechanism and only focus on seldom a part of content every time, and pass through
Accurate understanding to its meaning is realized to the concern repeatedly of important content.
Summary of the invention
The object of the present invention is to provide a kind of natural language semantic matching methods that sequence dynamic is read, and can pass through distich
The dynamic of subsequence, which is read, realizes the subsemantic accurate understanding of distich and expression, and then realizes the standard to natural language semantic matches
Really judgement.
The purpose of the present invention is what is be achieved through the following technical solutions:
A kind of natural language semantic matching method that sequence dynamic is read, comprising:
Semantic modeling is carried out to each word of natural language sentences centering, obtains the corresponding word grade of each natural language sentences
Other semantic expressiveness vector;
Vector is indicated according to each word level semantics, obtains corresponding sentence semantics expression vector by stacking neural network
And consider complementary hidden layer expression vector between word;
Vector is indicated using sentence semantics and considers complementary hidden layer expression vector progress sentence between word
Semantic dynamic understands that the dynamic for obtaining corresponding sentence, which understands, indicates vector;
The sentence semantics of natural language sentences pair are indicated that the dynamic of vector and sentence understands expression vector, are respectively carried out whole
It closes, and realizes that the semantic relation of natural language sentences pair is classified according to integrated results.
As seen from the above technical solution provided by the invention, for natural language sentences, the attention of the mankind is used for reference
Mechanism makes full use of the dynamic of distich subsequence to read, and the accurate selection to heavy duty word in sentence may be implemented and repeat to understand,
It is final to realize the matched accurate judgement of sentence semantics to realize to sentence semantics High Efficiency Modeling and characterization.
Detailed description of the invention
In order to illustrate the technical solution of the embodiments of the present invention more clearly, required use in being described below to embodiment
Attached drawing be briefly described, it should be apparent that, drawings in the following description are only some embodiments of the invention, for this
For the those of ordinary skill in field, without creative efforts, it can also be obtained according to these attached drawings other
Attached drawing.
Fig. 1 is a kind of process of natural language semantic matching method that sequence dynamic is read provided in an embodiment of the present invention
Figure.
Specific embodiment
With reference to the attached drawing in the embodiment of the present invention, technical solution in the embodiment of the present invention carries out clear, complete
Ground description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.Based on this
The embodiment of invention, every other implementation obtained by those of ordinary skill in the art without making creative efforts
Example, belongs to protection scope of the present invention.
The embodiment of the present invention provides a kind of natural language semantic matching method that sequence dynamic is read, as shown in Figure 1, it is led
Include:
Step 11 carries out semantic modeling to each word of natural language sentences centering, obtains each natural language sentences pair
The word level semantics answered indicate vector.
The preferred embodiment of this step is as follows:
1) natural language sentences pair are indicated using unified mathematical form: the natural language sentences to include two from
Right language sentence, a natural language sentences are denoted asIndicate it by laThe text of a word composition, separately
One natural language sentences is denoted asIndicate it by lbThe text of a word composition;Wherein,It is right
That answers respectively indicates natural language sentences saIn i-th of word, natural language sentences sbIn j-th of word.
2) natural language sentences are to saWith sbIn all words constitute a dictionary V, size lvIt indicates;Natural language
Say sentence to saAnd sbEach of word all use one solely hot vector (one-hot vector) indicate that vector length is word
The size of allusion quotation V, only its corresponding index position in dictionary V is 1 in only hot vector of each word, other are 0;Herein
On the basis of, the character representation of each word namely the word justice table of pre-training are obtained using the good term vector matrix E of pre-training
Show:
Wherein,Corresponding to natural language sentences saIn i-th of word, natural language sentences sbIn j-th it is single
The semantic expressiveness of the pre-training of word;
3) natural language sentences are assumed to being English text, then by the dictionary of one character set of all English alphabet compositions
Vc, size 26;Each of word letter indicates that vector length is dictionary V with an only hot vectorcSize, each
Only have it in dictionary V in only hot vector of lettercIn corresponding index position be 1, other are 0;On this basis, using one
Dimension convolution handles the alphabetical sequence of word respectively, and different convolution kernels (unigram, bigram, trigram) specifically can be used
Sentence is handled, is then operated using maximum pondization, to finally obtain each other semantic expressiveness of word character level:
Wherein, EcIndicate that the vector representing matrix of the character of needs training, Conv1D indicate one-dimensional convolution operation,
Maxpooling indicates maximum pondization operation,Corresponding expression natural language sentences saIn i-th of word i-thcIt is a
Letter only hotlist show, natural language sentences sbIn j-th of word jthcOnly hotlist of a letter shows;
4) each word is more fully indicated in order to more acurrate, by the character representation of the word obtained in the pre-training and corresponding
The other semantic expressiveness of word character level be stitched together, then using two layers high speed network (Highway network) integrate
These information, to finally obtain the semantic expressiveness vector of each word in natural language sentences:
Wherein, Highway () indicates high speed network structure, ai、bjCorresponding expression natural language sentences saI-th single
Semantic expressiveness vector, the natural language sentences s of wordbIn j-th of word semantic expressiveness vector.
Step 12 indicates vector according to each word level semantics, obtains corresponding sentence semantics by stacking neural network
It indicates vector and considers complementary hidden layer expression vector between word.
The preferred embodiment of this step is as follows:
1) mankind can be by reading understanding of the sentence intensification to the sentence semantics, therefore, in order to more comprehensively repeatedly
Modeling sentence give information, using stack Recognition with Recurrent Neural Network (Stack Recurrent Neural Network,
Stack-RNN) entire natural language sentences are modeled, obtain the implicit shape of each word in each natural language sentences
State sequence: it uses door recirculating network (GRU) and is used as basic unit, for the input x of t momentt, the hidden state h of GRUt
It updates as follows:
Z=σ (xtUz+ht-1Wz)
R=σ (xtUr+ht-1Wr)
Wherein, z, r, cmIt is update door, the resetting door, memory unit of GRU respectively;UzWith WzFor update door parameter matrix,
UrWith WrFor the parameter matrix for resetting door, UhWith WhFor the parameter matrix of memory unit,Indicate dot product;xtIndicate natural language sentence
Sub- saOr sbIn t-th of word semantic expressiveness vector;σ indicates Sigmoid activation operation.
On this basis, the subsemantic repeat reading of distich and reason are realized by the GRU of stacked multilayer (i.e. stack-GRU)
Solution is more completely understood sentence semantics to realize.But with the intensification of the network number of plies, model can not retain all acquired
Information, while be also faced with gradient disappear or explosion (gradient vanish or explore) problem.
In order to avoid problem above, the input of each layer of GRU and hidden layer output are spliced together by the embodiment of the present invention, make
Input for next layer:
Wherein, GRUlIndicate l layers of GRU,Indicate t-th of hidden layer state of (l-1) layer GRU,Indicate (l-
1) t-th of input of layer GRU, symbol [,] indicate concatenation;It is all to guarantee that model can retain for operation in this way
Information, while avoiding the problem that gradient disappears or explodes to a certain extent.
Later, using the mutation of this stack-GRU to natural language sentences to reading and understanding repeatedly are carried out, thus more
The semantic expressiveness for comprehensively encoding each word in each sentence obtains interdepending between word in each natural language sentences
Hidden layer indicate that formula is as follows:
Wherein,Corresponding expression natural language sentences saIn i-th ' a word, natural language sentences sbMiddle jth '
A word sentence level semantic expressiveness,Indicate natural language sentences saIn from the semantic table of the 1st phrase rank
Show the set of the semantic expressiveness of i-th ' a phrase rank,Indicate natural language sentences sbIn from the 1st phrase rank
Semantic expressiveness to jth ' a phrase rank semantic expressiveness set;
2) what is obtained above is complementary hidden layer expression between word in each sentence, and the semantic expressiveness of each word
The semantic expressiveness influence degree of entire sentence is different, and attention mechanism can help model to select and semantic expressiveness pass
The highest additional information of connection degree.
In order to guarantee the subsemantic accurate understanding of distich and expression, in the embodiment of the present invention, using from attention mechanism
(self-attention) semantic expressiveness of each word is obtained to the weighing factor of the semantic expressiveness of final sentence, and uses this
A little weights do weighted sum to the hidden layer state expression of all words, to obtain sentence semantics expression;
αa=ωT tanh(WAa+b)
Wherein, ω, W are from the weight in the calculating of attention mechanism, and b is to belong to from the biasing in the calculating of attention mechanism
Parameter during model training, αaIt indicates to natural language sentences saUse the weight obtained from after attention mechanism point
Cloth, i-th, a element of k ' be respectivelyhaIndicate natural language sentences saSemantic expressiveness vector;
Similarly, to natural language sentences sbUsing identical operation, natural language sentences s is obtainedbSemantic expressiveness vector
hb。
By this step, the embodiment of the present invention has obtained indicating by repeat reading and the sentence semantics of understanding, Yi Jikao
The hidden layer state for considering complementary each word between word indicates.
Step 13, indicate vector using sentence semantics and consider complementary hidden layer between word indicate vector into
The dynamic of row sentence semantics understands that the dynamic for obtaining corresponding sentence, which understands, indicates vector.
The preferred embodiment of this step is as follows:
1) previously mentioned, what the mankind can pay close attention to when understanding sentence semantics according to the information seen and the selection of desired information
Content, some words will not be read, and other word can be repeated reading.Meanwhile one is only focused on very in each concern
Small range.Therefore, the embodiment of the present invention proposes to select a most important word at each moment using a selection function,
And the most important word selected is handled by GRU, corresponding hidden layer state is obtained, continues to use selection function on this basis
The most important word of subsequent time is selected, is handled using GRU, obtains the hidden layer state of subsequent time, and repeat the process, directly
To reaching maximum dynamic sequence reading length.The last one hidden layer state will be understood by the dynamic as sentence indicates vector.By
Input of GRU is uncertain during this, and the information for needing to be grasped before basis calculates current input content, because
This process is referred to as dynamic and reads, and the sequence which goes out is referred to as dynamic and reads sequence:
Wherein, F indicates selection function,It indicates to correspond to natural language sentences saT-1 moment dynamic read sequence
Hidden layer state, hbIndicate natural language sentences sbSemantic expressiveness vector because the model be directed to sentence semantics matching, because
This is in processing natural language sentences saWhen need natural language sentences sbSemantic expressiveness vector hbAs additional supplemental information
It takes into account, lTIt indicates that dynamic reads the length of sequence, is previously set, vaIndicate natural language sentences saDynamic reason
Solution indicates vector;
Similarly, to natural language sentences sbUsing identical operation, natural language sentences s is obtainedbDynamic understand indicate to
Measure vb。
As previously mentioned, consider that attention mechanism can help model to select and semantic expressiveness correlation degree highest one
A word or several words can select most important word at each moment to realize, in the embodiment of the present invention, use is another
Kind attention mechanism selects the most important word of t moment
Wherein, ωd,Wd,Ud,MdIt indicates the weight in the calculating of attention mechanism, belongs to the parameter during model training,Indicate the semantic expressiveness of each word to the weighing factor distribution vector of the semantic expressiveness of final sentence,
The corresponding index value of the maximum value of weighing factor is selected in expression,Expression one is all 1 row vector, due toIt is a selection operation, is that gradient is not opened to guarantee that entire model can be led by softmax function
Above formula is modified as follow form by hair, the embodiment of the present invention:
Wherein, β is an arbitrarily large positive integer, it is contemplated that the characteristic of softmax function, arbitrarily large multiplied by one
After positive integer, the corresponding weight in more important position more levels off to 1, other weights more level off to 0.Pass through this kind of mode, this hair
Bright embodiment realizes guidable most important selected ci poem extract operation.
The sentence semantics of natural language sentences pair are indicated that the dynamic of vector and sentence understands expression vector by step 14, respectively
Realize that the semantic relation of natural language sentences pair is classified from being integrated, and according to integrated results.
The preferred embodiment of this step is as follows:
1) it in the embodiment of the present invention, is indicated using the sentence semantics that didactic method integrates natural language sentences pair respectively
The dynamic of vector and sentence, which understands, indicates that vector specifically can choose expression vector dot, subtract each other, and the operations such as splicing will
These characterization vectors integrate, and obtaining the sentence semantics between natural language sentences pair indicates the dynamic reason of vector h and sentence
Solution indicates vector v, is then found out under conditions of given different aspect information by multi-layer perception (MLP) (MLP), natural language sentence
The probability of semantic relation of the son between, the above process indicate are as follows:
H=(ha,hb,hb⊙ha,hb-ha),
V=(va,vb,vb⊙va,vb-va),
ph=MLP1(h),
pv=MLP1(v),
Wherein, ⊙ indicates dot product ,-indicate to subtract each other, () indicates concatenation;phIt indicates to utilize the nature after integration
Sentence semantics between language sentence pair indicate the semantic relation probability for the natural language sentences pair that vector h is calculated;pvTable
Show and understands the natural language sentence for indicating that vector v is calculated using the dynamic of the sentence between the natural language sentences pair after integration
The semantic relation probability of son pair.
MLP is a three-decker, defeated comprising two layers of full articulamentum and ReLu activation primitive and one layer of softmax
Layer out.In this layer, concatenation can retain the characterizing semantics information of sentence, available two sentences of dot product to greatest extent
Similarity information between son, the available characterizing semantics of phase reducing are different degrees of in each dimension.
2) sentence semantics more fully and are accurately understood in order to realize, by merging phWith pvAnd again to natural language
Semantic relation between speech sentence is classified: calculating p by linear transformationhWith pvShared weight, is then weighted and asks
With final semantic relation probability is acquired finally by another multi-layer perception (MLP) MLP:
P(y|(sa,sb))=MLP2(αhph+αvpv)
Wherein, ωh,ωvIt is phWith pvWeighting parameter in shared weight computations, bh,bvFor corresponding offset parameter,
σ indicates sigmoid function, αh、αvCorresponding expression ph、pvShared weight;P(y|(sa,sb)) indicate natural language sentences pair
saWith sbBetween semantic relation probability distribution.
Above scheme provided in an embodiment of the present invention, not only by stacking readding repeatedly for Recognition with Recurrent Neural Network distich subsequence
It reads, realizes the subsemantic more thorough understanding of distich;And sequential structure is read to the essence of primary word in sentence by dynamic
It really selects and reads repeatedly, to realize the subsemantic more fully accurate understanding of distich and expression, and then High Efficiency Modeling two
Semantic interaction between a sentence, the semantic deduction relationship between two sentences of final accurate judgement, while additionally providing one kind
Accurate sentence semantics characterizing method compensates for existing method insufficient present on sentence semantics expression.
Through the above description of the embodiments, those skilled in the art can be understood that above-described embodiment can
The mode of necessary general hardware platform can also be added to realize by software by software realization.Based on this understanding,
The technical solution of above-described embodiment can be embodied in the form of software products, which can store non-easy at one
In the property lost storage medium (can be CD-ROM, USB flash disk, mobile hard disk etc.), including some instructions are with so that a computer is set
Standby (can be personal computer, server or the network equipment etc.) executes method described in each embodiment of the present invention.
The foregoing is only a preferred embodiment of the present invention, but scope of protection of the present invention is not limited thereto,
Within the technical scope of the present disclosure, any changes or substitutions that can be easily thought of by anyone skilled in the art,
It should be covered by the protection scope of the present invention.Therefore, protection scope of the present invention should be with the protection model of claims
Subject to enclosing.
Claims (6)
1. a kind of natural language semantic matching method that sequence dynamic is read characterized by comprising
Semantic modeling is carried out to each word of natural language sentences centering, obtains the corresponding word rank language of each natural language sentences
Justice indicates vector;
Indicate vector according to each word level semantics, by stack neural network obtain corresponding sentence semantics indicate vector and
Considering complementary hidden layer between word indicates vector;
Vector is indicated using sentence semantics and considers complementary hidden layer expression vector progress sentence semantics between word
Dynamic understand, obtain corresponding sentence dynamic understand indicate vector;
The sentence semantics of natural language sentences pair are indicated that the dynamic of vector and sentence understands expression vector, are respectively integrated,
And realize that the semantic relation of natural language sentences pair is classified according to integrated results.
2. a kind of natural language semantic matching method that sequence dynamic is read according to claim 1, which is characterized in that institute
It states and semantic modeling is carried out to each word of natural language sentences centering, obtain the corresponding word level semantics of each natural language sentences
Indicate vector the step of include:
To including two natural language sentences, a natural language sentences are denoted as the natural language sentencesIndicate it by laThe text of a word composition, another natural language sentences are denoted asIndicate it by lbThe text of a word composition;Wherein,It is corresponding to respectively indicate natural language
Sentence saIn i-th of word, natural language sentences sbIn j-th of word;
Natural language sentences are to saWith sbIn all words constitute a dictionary V, size lvIt indicates;Natural language sentences
To saAnd sbEach of word all indicated with an only hot vector, vector length be dictionary V size, each word it is only
Only its corresponding index position in dictionary V is 1 in hot vector, other are 0;On this basis, good using pre-training
Term vector matrix E obtains the character representation of each word namely the word semantic expressiveness of pre-training:
Wherein,Corresponding to natural language sentences saIn i-th of word, natural language sentences sbIn j-th word
The semantic expressiveness of pre-training;
Assuming that all English alphabets are then formed the dictionary V of a character set to for English text by natural language sentencesc, big
Small is 26;Each of word letter indicates that vector length is dictionary V with an only hot vectorcSize, it is each letter
Only have it in dictionary V in only hot vectorcIn corresponding index position be 1, other are 0;On this basis, using one-dimensional convolution
The alphabetical sequence of word is handled respectively, is then operated using maximum pondization, to finally obtain each other language of word character level
Justice indicates:
Wherein, EcIndicate that the vector representing matrix of the character of needs training, Conv1D indicate one-dimensional convolution operation, Maxpooling
Indicate maximum pondization operation,Corresponding expression natural language sentences saIn i-th of word i-thcOnly heat of a letter
It indicates, natural language sentences sbIn j-th of word jthcOnly hotlist of a letter shows;
The character representation of the word obtained in the pre-training is stitched together with the corresponding other semantic expressiveness of word character level again,
Then these information are integrated using two layers of high speed network, to finally obtain the semantic table of each word in natural language sentences
Show vector:
Wherein, Highway () indicates high speed network structure, ai、bjCorresponding expression natural language sentences saI-th word
Semantic expressiveness vector, natural language sentences sbIn j-th of word semantic expressiveness vector.
3. a kind of natural language semantic matching method that sequence dynamic is read according to claim 2, which is characterized in that institute
Stating indicates vector according to each word level semantics, obtains corresponding sentence semantics expression vector by stacking neural network and examines
Having considered the step of complementary hidden layer indicates vector between word includes:
Entire natural language sentences are modeled using Recognition with Recurrent Neural Network is stacked, are obtained every in each natural language sentences
The hidden state sequence of a word: door recirculating network GRU is used as basic unit, for the input x of t momentt,
The hidden state h of GRUtIt updates as follows:
Z=σ (xtUz+ht-1Wz)
R=σ (xtUr+ht-1Wr)
Wherein, z, r, cmIt is update door, the resetting door, memory unit of GRU respectively;UzWith WzFor the parameter matrix for updating door, UrWith
WrFor the parameter matrix for resetting door, UhWith WhFor the parameter matrix of memory unit,Indicate dot product;xtIndicate natural language sentences sa
Or sbIn t-th of word semantic expressiveness vector;σ indicates Sigmoid activation operation;
On this basis, by stacked multilayer GRU, i.e. stack-GRU, the input of each layer of GRU and hidden layer output are spliced to
Together, as next layer of input:
Wherein, GRUlIndicate l layers of GRU,Indicate t-th of hidden layer state of (l-1) layer GRU,Indicate (l-1) layer
T-th of input of GRU, symbol [,] indicate concatenation;
Using stack-GRU to natural language sentences to reading and understanding repeatedly are carried out, obtain single in each natural language sentences
Complementary hidden layer indicates between word, and formula is as follows:
Wherein,Corresponding expression natural language sentences saIn i-th ' a word, natural language sentences sbMiddle jth ' a list
Word sentence level semantic expressiveness,Indicate natural language sentences saIn from the semantic expressiveness of the 1st phrase rank to
The set of the semantic expressiveness of i-th ' a phrase rank,Indicate natural language sentences sbIn from the language of the 1st phrase rank
Justice indicates the set of the semantic expressiveness to jth ' a phrase rank;
Using obtaining the semantic expressiveness of each word from attention mechanism to the weighing factor of the semantic expressiveness of final sentence, and make
It indicates to do weighted sum with hidden layer state of these weights to all words, to obtain sentence semantics expression;
Wherein, ω, W are from the weight in the calculating of attention mechanism, and b is to belong to mould from the biasing in the calculating of attention mechanism
Parameter in type training process, αaIt indicates to natural language sentences saIt is distributed using the weight obtained after attention mechanism, haTable
Show natural language sentences saSemantic expressiveness vector;
Similarly, to natural language sentences sbUsing identical operation, natural language sentences s is obtainedbSemantic expressiveness vector hb。
4. a kind of natural language semantic matching method that sequence dynamic is read according to claim 3, which is characterized in that institute
It states using sentence semantics expression vector and considers complementary hidden layer expression vector progress sentence semantics between word
Dynamic understands that the dynamic for obtaining corresponding sentence understands that the step of indicating vector includes:
Select a most important word at each moment using a selection function, and handled by GRU select it is most heavy
The word wanted obtains corresponding hidden layer state, continues to use selection function on this basis and selects the most important word of subsequent time, benefit
It is handled with GRU, obtains the hidden layer state of subsequent time, and repeat the process, until reaching the maximum dynamic sequence of setting
Reading length, finally understanding the last one hidden layer state as the dynamic of sentence indicates vector:
Wherein, F indicates selection function,It indicates to correspond to natural language sentences saThe t-1 moment dynamic read sequence hidden layer
State, hbIndicate natural language sentences sbSemantic expressiveness vector, lTIndicate that dynamic reads the length of sequence, vaIndicate nature language
Say sentence saDynamic understand indicate vector;
Similarly, to natural language sentences sbUsing identical operation, natural language sentences s is obtainedbDynamic understand indicate vector
vb。
5. a kind of natural language semantic matching method that sequence dynamic is read according to claim 4, which is characterized in that
The most important word of t moment is selected using another attention mechanism
Wherein, ωd,Wd,Ud,MdWeight in the calculating of attention mechanism, belongs to the parameter during model training,Indicate every
The semantic expressiveness of a word to the weighing factor distribution vector of the semantic expressiveness of final sentence,Expression is selected
The corresponding index value of the maximum value of weighing factor,Indicate one be all 1 row vector, β is one arbitrarily large just whole
Number.
6. a kind of natural language semantic matching method that sequence dynamic is read according to claim 4, which is characterized in that will
The sentence semantics of natural language sentences pair indicate that the dynamic of vector and sentence understands expression vector, are respectively integrated, and according to
Integrated results realize that the step of semantic relation classification of natural language sentences pair includes:
Indicate that the dynamic of vector and sentence understands using the sentence semantics that didactic method integrates natural language sentences pair respectively
Indicate vector, obtaining the sentence semantics between natural language sentences pair indicates that the dynamic of vector sum sentence understands expression vector, so
It is found out under conditions of given different aspect information by multi-layer perception (MLP) MLP afterwards, the semantic pass between natural language sentences pair
The probability of system, the above process indicate are as follows:
H=(ha,hb,hb⊙ha,hb-ha),
V=(va,vb,vb⊙va,vb-va),
ph=MLP1(h),
pv=MLP1(v),
Wherein, ⊙ indicates dot product ,-indicate to subtract each other, () indicates concatenation;phIt indicates to utilize the natural language sentence after integration
Sentence semantics of the son between indicate the semantic relation probability for the natural language sentences pair that vector h is calculated;pvIt indicates to utilize
The dynamic of the sentence between natural language sentences pair after integration understands the natural language sentences pair for indicating that vector v is calculated
Semantic relation probability;
By merging phWith pvAnd classify again to the semantic relation between natural language sentences: being calculated by linear transformation
P outhWith pvThen shared weight is weighted summation, acquire final language finally by another multi-layer perception (MLP) MLP
Adopted relationship probability:
P(y|(sa,sb))=MLP2(αhph+αvpv)
Wherein, ωh,ωvIt is phWith pvWeighting parameter in shared weight computations, bh,bvFor corresponding offset parameter, σ table
Show sigmoid function, αh、αvCorresponding expression ph、pvShared weight;P(y|(sa,sb)) indicate natural language sentences to saWith
sbBetween semantic relation probability distribution.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910228242.9A CN109948163B (en) | 2019-03-25 | 2019-03-25 | Natural language semantic matching method for dynamic sequence reading |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910228242.9A CN109948163B (en) | 2019-03-25 | 2019-03-25 | Natural language semantic matching method for dynamic sequence reading |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109948163A true CN109948163A (en) | 2019-06-28 |
CN109948163B CN109948163B (en) | 2022-10-28 |
Family
ID=67010836
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910228242.9A Active CN109948163B (en) | 2019-03-25 | 2019-03-25 | Natural language semantic matching method for dynamic sequence reading |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109948163B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110705311A (en) * | 2019-09-27 | 2020-01-17 | 安徽咪鼠科技有限公司 | Semantic understanding accuracy improving method, device and system applied to intelligent voice mouse and storage medium |
CN110765240A (en) * | 2019-10-31 | 2020-02-07 | 中国科学技术大学 | Semantic matching evaluation method for multiple related sentence pairs |
CN111859909A (en) * | 2020-07-10 | 2020-10-30 | 山西大学 | Semantic scene consistency recognition reading robot |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2018033030A1 (en) * | 2016-08-19 | 2018-02-22 | 中兴通讯股份有限公司 | Natural language library generation method and device |
CN109214006A (en) * | 2018-09-18 | 2019-01-15 | 中国科学技术大学 | The natural language inference method that the hierarchical semantic of image enhancement indicates |
CN109344404A (en) * | 2018-09-21 | 2019-02-15 | 中国科学技术大学 | The dual attention natural language inference method of context aware |
-
2019
- 2019-03-25 CN CN201910228242.9A patent/CN109948163B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2018033030A1 (en) * | 2016-08-19 | 2018-02-22 | 中兴通讯股份有限公司 | Natural language library generation method and device |
CN109214006A (en) * | 2018-09-18 | 2019-01-15 | 中国科学技术大学 | The natural language inference method that the hierarchical semantic of image enhancement indicates |
CN109344404A (en) * | 2018-09-21 | 2019-02-15 | 中国科学技术大学 | The dual attention natural language inference method of context aware |
Non-Patent Citations (2)
Title |
---|
徐丽丽等: "面向机器阅读理解的语句填补答案选择方法", 《计算机工程》 * |
黄江平等: "基于卷积网络的句子语义相似性模型", 《华南理工大学学报(自然科学版)》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110705311A (en) * | 2019-09-27 | 2020-01-17 | 安徽咪鼠科技有限公司 | Semantic understanding accuracy improving method, device and system applied to intelligent voice mouse and storage medium |
CN110705311B (en) * | 2019-09-27 | 2022-11-25 | 安徽咪鼠科技有限公司 | Semantic understanding accuracy improving method, device and system applied to intelligent voice mouse and storage medium |
CN110765240A (en) * | 2019-10-31 | 2020-02-07 | 中国科学技术大学 | Semantic matching evaluation method for multiple related sentence pairs |
CN110765240B (en) * | 2019-10-31 | 2023-06-20 | 中国科学技术大学 | Semantic matching evaluation method for multi-phase sentence pairs |
CN111859909A (en) * | 2020-07-10 | 2020-10-30 | 山西大学 | Semantic scene consistency recognition reading robot |
CN111859909B (en) * | 2020-07-10 | 2022-05-31 | 山西大学 | Semantic scene consistency recognition reading robot |
Also Published As
Publication number | Publication date |
---|---|
CN109948163B (en) | 2022-10-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110188358B (en) | Training method and device for natural language processing model | |
CN109992779B (en) | Emotion analysis method, device, equipment and storage medium based on CNN | |
CN109657246B (en) | Method for establishing extraction type machine reading understanding model based on deep learning | |
Gallant et al. | Representing objects, relations, and sequences | |
CN110096711B (en) | Natural language semantic matching method for sequence global attention and local dynamic attention | |
CN108830287A (en) | The Chinese image, semantic of Inception network integration multilayer GRU based on residual error connection describes method | |
WO2020247616A1 (en) | Linguistically rich cross-lingual text event embeddings | |
CN110222163A (en) | A kind of intelligent answer method and system merging CNN and two-way LSTM | |
Li et al. | A method of emotional analysis of movie based on convolution neural network and bi-directional LSTM RNN | |
CN109241377A (en) | A kind of text document representation method and device based on the enhancing of deep learning topic information | |
CN112232087B (en) | Specific aspect emotion analysis method of multi-granularity attention model based on Transformer | |
CN109214006A (en) | The natural language inference method that the hierarchical semantic of image enhancement indicates | |
CN108763535A (en) | Information acquisition method and device | |
CN109948163A (en) | The natural language semantic matching method that sequence dynamic is read | |
CN112800190A (en) | Intent recognition and slot value filling joint prediction method based on Bert model | |
CN111753088A (en) | Method for processing natural language information | |
CN112232053A (en) | Text similarity calculation system, method and storage medium based on multi-keyword pair matching | |
CN112000778A (en) | Natural language processing method, device and system based on semantic recognition | |
CN112988970A (en) | Text matching algorithm serving intelligent question-answering system | |
CN114254645A (en) | Artificial intelligence auxiliary writing system | |
CN111914553A (en) | Financial information negative subject judgment method based on machine learning | |
Li et al. | Multimodal fusion with co-attention mechanism | |
Rodzin et al. | Deep learning techniques for natural language processing | |
Tekir et al. | Deep learning: Exemplar studies in natural language processing and computer vision | |
CN111581365B (en) | Predicate extraction method |
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 |