CN106933785A - A kind of abstraction generating method based on recurrent neural network - Google Patents
A kind of abstraction generating method based on recurrent neural network Download PDFInfo
- Publication number
- CN106933785A CN106933785A CN201710099638.9A CN201710099638A CN106933785A CN 106933785 A CN106933785 A CN 106933785A CN 201710099638 A CN201710099638 A CN 201710099638A CN 106933785 A CN106933785 A CN 106933785A
- Authority
- CN
- China
- Prior art keywords
- state vector
- recurrent neural
- neural network
- decoder
- method based
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 48
- 230000000306 recurrent effect Effects 0.000 title claims abstract description 37
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 36
- 239000013598 vector Substances 0.000 claims abstract description 126
- 230000008569 process Effects 0.000 claims description 11
- 230000006870 function Effects 0.000 claims description 10
- 239000010410 layer Substances 0.000 claims description 10
- 238000010276 construction Methods 0.000 claims description 2
- 230000001537 neural effect Effects 0.000 claims description 2
- 239000002356 single layer Substances 0.000 claims description 2
- 230000007246 mechanism Effects 0.000 description 14
- 230000000694 effects Effects 0.000 description 8
- 238000010586 diagram Methods 0.000 description 4
- 238000010801 machine learning Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- JEIPFZHSYJVQDO-UHFFFAOYSA-N iron(III) oxide Inorganic materials O=[Fe]O[Fe]=O JEIPFZHSYJVQDO-UHFFFAOYSA-N 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 238000012549 training Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 230000007935 neutral effect Effects 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/10—Text processing
- G06F40/12—Use of codes for handling textual entities
- G06F40/126—Character encoding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/10—Text processing
- G06F40/12—Use of codes for handling textual entities
- G06F40/151—Transformation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Compression, Expansion, Code Conversion, And Decoders (AREA)
Abstract
The present invention relates to a kind of abstraction generating method based on recurrent neural network, in current time t, by the state vector s of recurrent neural network decodertState vector with each moment of recurrent neural network encoder is contrasted, and is found out and state vector stThe most strong state vector H of relevance, then d state vector of utilization state vector H and the state vector H left and right sides be calculated state vector ct, utilization state vector ct, state vector stIt is calculated new state vector dt, then according to state vector dtDecoding obtains next word or word of output sequence, and wherein d is the integer more than or equal to 1.
Description
Technical field
The present invention relates to field of neural networks, more particularly, to a kind of summarization generation side based on recurrent neural network
Method.
Background technology
Summarization generation is a major issue of natural language processing field, and it mainly has two kinds of different forms:It is a kind of
It is the purport for generating source document, another kind is the title for generating source document.The former is general long, potentially include dozens of word or
Word, and the latter is relatively short, general only ten words or so.Summary is that the height of source document is summarized, and it must be simple and clear
Expression source document the meaning.Traditional abstraction generating method can be divided into three steps:First, according to certain standard (such as participle)
Source document is divided into many small pieces;2nd, according to each segment weight (such as tf-idf), therefrom select weight ratio compared with
Those big segments;3rd, the larger segment of those weights is combined into by new sentence according to certain algorithm, as plucking for source document
Will.
Prior art provides a kind of abstraction generating method for being based on recurrent neural network encoder and decoder, should
Method is actually machine-learning process of the sequence to sequence, and its input can be a sentence, a paragraph or
One article, its output is the purport or title of corresponding input, therefore input and output can be regarded as by word or
The time series of word composition.Compared with traditional abstraction generating method, this is a kind of abstract summarization generation process.Its basis is given
Fixed list entries, method by whole vocabulary, recursively make constituting a new sentence from front to back by search keyword
Be output sequence, i.e. summary info.
The effect of wherein recurrent neural network encoder be using recurrent neural network will be given list entries conversion or
Person's mapping turns into an intermediate expression, and the paragraph or article that will be input into are converted into a vector expression H.Assuming that input
Sequence is:X={ x1,x2,…,xn, wherein n represents the length of list entries.As shown in figure 1, the effect of encoder can be used to
Lower expression formula is represented:
ht=f (xt,ht-1)
Wherein, xtRepresent t-th corresponding vector of element, h in list entriestRepresent the state vector of t encoder, f
WithRepresent nonlinear mapping function.H represents the vector expression of list entries, generally takes H=hn, i.e., with passing
Return neutral net encoder last moment state vector as list entries intermediate vector expression formula.Eos is a spy
Different mark, represents the termination of list entries, and encoder work end and the beginning of decoder functions.
Accordingly, the effect of recurrent neural network decoder is the intermediate vector expression formula H next life using encoder generation
Into output sequence.Assuming that output sequence is:Y={ y1,y2,…,ym, wherein m represents the length of output sequence.Should be noted
It is that decoder is not disposably to generate whole output sequence, but is generated according to each moment of order from front to back and exported
One word or a word of order, untill the whole output sequence of generation.As shown in Fig. 2 the effect of decoder can be used
Following formula is represented:
P(yt|y1..., yt-1, H) and=g (st,H)
st=f (yt-1, st-1)
Wherein, P (Y │ X) represents the probability for obtaining being input into Y according to input X;ytRepresent the decoding of t in output sequence
The word or word for going out, stRepresent the state vector of t decoder.F and g represent non-linear transfer function, and g is used here
Softmax functions.
In such scheme, the abstraction generating method based on recurrent neural network encoder and decoder has a deficiency, i.e.,
Output sequence generation only it is relevant with the state vector of the last moment of encoder, but with encoder in other state vectors without
Close.When recurrent neural network length increase when, recurrent neural network extract characteristic vector often with list entries after
The state relation in face increases, but is reduced with the state relation before list entries, and this may cause the decay of information.
Therefore only being decoded according to the final state of encoder can cause the relevance between output sequence and list entries to die down.
The content of the invention
The present invention can cause the defect of information attenuation to solve above prior art during decoding, there is provided a kind of
Abstraction generating method based on recurrent neural network
To realize above goal of the invention, the technical scheme of use is:
A kind of abstraction generating method based on recurrent neural network, in current time t, by recurrent neural network decoder
State vector stState vector with each moment of recurrent neural network encoder is contrasted, and is found out and state vector stClose
The most strong state vector H of connection property, then d state vector of utilization state vector H and the state vector H left and right sides be calculated
State vector ct, utilization state vector ct, state vector stObtain new state vector dt, then according to state vector dtDecode
To next word or word of output sequence, wherein d is the integer more than or equal to 1.
In such scheme, when next word or word of output sequence to be generated, the method that the present invention is provided is simultaneously
Do not decoded with the current state vector of decoder directly;But the current state vector s of decodertWith each moment of encoder
State vector contrasted, find out and stThe state vector H of most like encoder, then by H and its both sides several
State vector is calculated a state vector ct, and utilization state vector ct, state vector stIt is calculated new state vector
dt, it has reacted the word currently to be generated or word should be proportionate with it, i.e., it is the next word or word that will be generated
Alignment information.Finally according to state vector dtNext word or word of output sequence are obtained to decode.By finding H and profit
State vector c is generated with Ht, and utilization state vector ct, state vector stIt is calculated new state vector dt, finally using shape
State vector dtDecoding obtains next word or word of output sequence so that method is solved between output sequence and list entries
Alignment relation, improve the quality and efficiency of output sequence, and make relevance between output sequence and list entries so as to
Level higher can be maintained at.
Preferably, it is described to find out and state vector stThe detailed process of the most strong state vector H of relevance is as follows:
Wherein ptPositions of the state vector H in recurrent neural network coder state sequence vector is represented, n represents input
The length of sequence,wpThe parameter for needing study is represented, sigmoid functions are specifically expressed as follows:
Sigmoid (x)=1/ (1+e-x)
Tanh function representations are as follows:
Preferably, it is described to be calculated state vector ctDetailed process it is as follows:
Wherein, αtiRepresent weight, hiRepresent the state vector of recurrent neural network encoder.
Preferably, the αtiTo ask for process as follows:
Wherein etiRepresent the relevance weight of state vector in state vector and encoder in decoder:
eti=si*hi。
Preferably, the state vector dtCalculating process it is as follows:
dt=sigmoid (ct*st)
Wherein, sigmoid functions are specifically expressed as follows:
Sigmoid (x)=1/ (1+e-x)。
In such scheme, state vector dtIt is in fact a kind of strategy of local optimum for the treatment of alignment relation, output
Sequence and the alignment relation of list entries should be overall related rather than should be local correlation, when this means that current
The alignment information at quarter not only has an impact to the decoding at current time, and should can also have an impact to follow-up decoding.Therefore, originally
The alignment information d of the method current time of offer is providedtAs in next state of extra information input to decoder.
Preferably, if recurrent neural network encoder is sandwich construction with decoder, by state vector ctAnd/or dtDirectly
Connect and be input to the ground floor of decoder or last layer and be applied in the decoding of decoder subsequent time;If recurrent neural net
Network encoder is single layer structure with decoder, then by state vector ctAnd/or dtIt is directly inputted to decoder.
Compared with prior art, the beneficial effects of the invention are as follows:
When next word or word of output sequence is generated, the method that the present invention is provided is not for the method for the present invention
Directly decoded with the current state vector of decoder;But the current state vector s of decodertWith each moment of encoder
State vector is contrasted, and is found out and stThe state vector H of most like encoder, then by H and several shapes of its both sides
State vector is calculated a state vector ct, and utilization state vector ct, state vector stIt is calculated new state vector
dt, it has reacted the word currently to be generated or word should be proportionate with it, i.e., it is the next word or word that will be generated
Alignment information.Finally according to state vector dtNext word or word of output sequence are obtained to decode.By finding H and profit
State vector c is generated with Ht, and utilization state vector ct, state vector stIt is calculated new state vector dt, finally using shape
State vector dtDecoding obtains next word or word of output sequence so that method is solved between output sequence and list entries
Alignment relation, improve the quality and efficiency of output sequence, and make relevance between output sequence and list entries so as to
Level higher can be maintained at.
Brief description of the drawings
Fig. 1 is the effect schematic diagram of encoder.
Fig. 2 is the effect schematic diagram of decoder.
Fig. 3 is the effect schematic diagram of attention mechanism.
(a), (b), (c) of Fig. 4, (d) are respectively four kinds of effect schematic diagrames of difference feed mechanism.
Specific embodiment
Accompanying drawing being for illustration only property explanation, it is impossible to be interpreted as the limitation to this patent;
Below in conjunction with drawings and Examples, the present invention is further elaborated.
Embodiment 1
The primary object of the method that the present invention is provided is that increased attention mechanism, as shown in figure 3, its is specific
It is as follows:
In current time t, by the state vector s of recurrent neural network decodertIt is every with recurrent neural network encoder
The state vector at individual moment is contrasted, and is found out and state vector stThe most strong state vector H of relevance, then utilization state to
D state vector of amount H and the state vector H left and right sides is calculated state vector ct, utilization state vector ct, state vector
stObtain new state vector dt, then according to state vector dtDecoding obtains next word or word of output sequence, wherein d
It is the integer more than or equal to 1.
In such scheme, when next word or word of output sequence to be generated, the method that the present invention is provided is simultaneously
Do not decoded with the current state vector of decoder directly;But the current state vector s of decodertWith each moment of encoder
State vector contrasted, find out and stThe state vector H of most like encoder, then by H and its both sides several
State vector is calculated a state vector ct, and utilization state vector ct, state vector stIt is calculated new state vector
dt, it has reacted the word currently to be generated or word should be proportionate with it, i.e., it is the next word or word that will be generated
Alignment information.Finally according to state vector dtNext word or word of output sequence are obtained to decode.By finding H and profit
State vector c is generated with Ht, and utilization state vector ct, state vector stIt is calculated new state vector dt, finally using shape
State vector dtDecoding obtains next word or word of output sequence so that method is solved between output sequence and list entries
Alignment relation, improve the quality and efficiency of output sequence, and make relevance between output sequence and list entries so as to
Level higher can be maintained at.
Embodiment 2
Attention mechanism solves the alignment relation between output sequence and list entries to a certain extent, so that
Improve the quality and efficiency of output sequence.But attention mechanism is in fact one kind local most for the treatment of alignment relation
Excellent strategy.In order to further improve the quality of output sequence i.e. source document summary, the present embodiment has used a kind of new complete
The optimal strategy of office, i.e. feed mechanism.
There are two kinds of alignment informations in fact in attention mechanism, one kind is direct alignment information ct, another kind is
Indirectly alignment informationd t, they have all reacted position and the content that next word or word in output sequence should align.Output sequence
Row should be overall related rather than that should be local correlation to the alignment relation of list entries, and this means that current time
Alignment information not only have an impact to the decoding at current time, and should can also have an impact to follow-up decoding.Therefore, it can
Using the alignment information at current time as extra information input to next state of decoder in.
Can be individual layer, or multilayer based on recurrent neural network encoder and decoder, as shown in Figure 4.
The present embodiment employs two kinds of distribution mechanisms:One is the subsequent time which kind of alignment information is input to decoder, is directly alignd
Information or indirectly alignment information.Two is that last layer is also to which layer of the subsequent time of decoder by alignment information input
It is last layer.Based on both mechanism, four kinds of difference feed mechanism can be obtained, as shown in figure 4, they are respectively:Indirectly
Alignment information is input to last layer, and direct alignment information is input to last layer, and indirect alignment information is input to ground floor, directly
Connect alignment information and be input to ground floor.
Embodiment 3
The sequence of summarization generation essentially belongs to a kind of machine learning method to the learning method of sequence, therefore
Can be trained using the general training method of machine learning.In order to accelerate training speed, the present embodiment is using minimum lot size ladder
Descent algorithm (mini-batch gradient descent) is spent to train.Because the vocabulary of Chinese is very big, needed for decoding process
Want the time very long, in order to reduce the decoding time, the present embodiment uses word table, i.e., only use 4000 the most frequently used Chinese character conducts
Word table, for other words, is replaced using special mark.Tables 1 and 2 is made by using test data set to generation method
Test and evaluation
Table 1
Model | R-1 | R-2 | R-L | BLEU |
Multilayer RNN | 30.4 | 16.0 | 27.3 | 9.5 |
Multilayer RNN+attention | 32.2 | 17.8 | 28.8 | 11.4 |
Multilayer RNN+attention+feed1 | 32.2 | 17.7 | 28.7 | 11.3 |
Multilayer RNN+attention+feed2 | 33.1 | 18.3 | 29.5 | 12.0 |
Multilayer RNN+attention+feed3 | 30.1 | 16.7 | 27.6 | 10.3 |
Multilayer RNN+attention+feed4 | 31.1 | 17.1 | 27.9 | 10.9 |
Table 2
Model | Informedness | Grammer | Terseness |
Multilayer RNN+attention | 2.87 | 3.95 | 3.28 |
Multilayer RNN+attention+feed1 | 2.85 | 4.01 | 2.93 |
Multilayer RNN+attention+feed2 | 3.02 | 4.23 | 3.30 |
Multilayer RNN+attention+feed3 | 3.00 | 3.93 | 3.10 |
Multilayer RNN+attention+feed4 | 2.88 | 4.03 | 3.03 |
Manually | 3.88 | 4.42 | 3.80 |
Table 1 employs two kinds of machine index ROUGE and BLEU to be estimated generation method.ROUGE is from recall rate
Angle assess, and BLEU is angle from accurate rate assesses.From table 1 it follows that using second feed
Strategy, will be when directly alignment information be input to last layer of decoder, and the result of test behaves oneself best, more than only making
With the generation method of attention mechanism, this illustrates the validity of generation method.
Table 2 employs three-type-person's work index to assess generation method.Wherein informedness reflects the forgiven information content of summary,
The degree of grammer reflection summary grammatical, the succinct degree of terseness reflection summary, artificial model represents manually generated
Summary info.Can show that the conclusion similar with table 1, i.e. feed mechanism have exceeded attention mechanism according to table 2.Equally will
Last layer of this strategy that direct alignment information is input to decoder shows outstanding.
By two contrasts of form, it can be deduced that abstraction generating method has high efficiency, the conclusion of high accuracy, especially
When being using second feed mechanism.
Obviously, the above embodiment of the present invention is only intended to clearly illustrate example of the present invention, and is not right
The restriction of embodiments of the present invention.For those of ordinary skill in the field, may be used also on the basis of the above description
To make other changes in different forms.There is no need and unable to be exhaustive to all of implementation method.It is all this
Any modification, equivalent and improvement made within the spirit and principle of invention etc., should be included in the claims in the present invention
Protection domain within.
Claims (7)
1. a kind of abstraction generating method based on recurrent neural network, it is characterised in that:In current time t, by recurrent neural net
The state vector s of network decodertState vector with each moment of recurrent neural network encoder is contrasted, and is found out and shape
State vector stThe most strong state vector H of relevance, then d state of utilization state vector H and the state vector H left and right sides to
Amount is calculated state vector ct, utilization state vector ct, state vector stIt is calculated new state vector dt, then basis
State vector dtDecoding obtains next word or word of output sequence, and wherein d is the integer more than or equal to 1.
2. the abstraction generating method based on recurrent neural network according to claim 1, it is characterised in that:
It is described to find out and state vector stThe detailed process of the most strong state vector H of relevance is as follows:
Wherein ptPositions of the state vector H in recurrent neural network coder state sequence vector is represented, n represents list entries
Length,wpThe parameter for needing study is represented, sigmoid functions are specifically expressed as follows:
Sigmoid (x)=1/ (1+e-x)
Tanh function representations are as follows:
3. the abstraction generating method based on recurrent neural network according to claim 2, it is characterised in that:It is described to calculate
To state vector ctDetailed process it is as follows:
Wherein, αtiRepresent weight, hiRepresent the state vector of recurrent neural network encoder.
4. the abstraction generating method based on recurrent neural network according to claim 3, it is characterised in that:The αtiAsk
Take process as follows:
Wherein etiRepresent the relevance weight of state vector in state vector and encoder in decoder:
eti=si*hi。
5. the abstraction generating method based on recurrent neural network according to claim 1, it is characterised in that:
The state vector dtCalculating process it is as follows:
dt=sigmoid (ct*st)
Wherein, sigmoid functions are specifically expressed as follows:
Sigmoid (x)=1/ (1+e-x)。
6. the abstraction generating method based on recurrent neural network according to claim 1, it is characterised in that:By state vector
ctAnd/or dtIt is added in the decoder of decoder subsequent time.
7. the abstraction generating method based on recurrent neural network according to claim 6, it is characterised in that:If recurrent neural
Network encoder is sandwich construction with decoder, then by state vector ctAnd/or dtIt is directly inputted to the ground floor or most of decoder
Later layer is applied in the decoding of decoder subsequent time;If recurrent neural network encoder is single layer structure with decoder,
Then by state vector ctAnd/or dtIt is directly inputted to decoder.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710099638.9A CN106933785A (en) | 2017-02-23 | 2017-02-23 | A kind of abstraction generating method based on recurrent neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710099638.9A CN106933785A (en) | 2017-02-23 | 2017-02-23 | A kind of abstraction generating method based on recurrent neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106933785A true CN106933785A (en) | 2017-07-07 |
Family
ID=59423748
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710099638.9A Pending CN106933785A (en) | 2017-02-23 | 2017-02-23 | A kind of abstraction generating method based on recurrent neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106933785A (en) |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107766506A (en) * | 2017-10-20 | 2018-03-06 | 哈尔滨工业大学 | A kind of more wheel dialog model construction methods based on stratification notice mechanism |
CN108153864A (en) * | 2017-12-25 | 2018-06-12 | 北京牡丹电子集团有限责任公司数字电视技术中心 | Method based on neural network generation text snippet |
CN108319668A (en) * | 2018-01-23 | 2018-07-24 | 义语智能科技(上海)有限公司 | Generate the method and apparatus of text snippet |
CN108984524A (en) * | 2018-07-05 | 2018-12-11 | 北京理工大学 | A kind of title generation method based on variation neural network topic model |
CN109325110A (en) * | 2018-08-24 | 2019-02-12 | 广东外语外贸大学 | Indonesian documentation summary generation method, device, storage medium and terminal device |
CN109947930A (en) * | 2019-03-12 | 2019-06-28 | 上海秘塔网络科技有限公司 | Abstraction generating method, device, terminal and computer readable storage medium |
CN109948162A (en) * | 2019-03-25 | 2019-06-28 | 北京理工大学 | The production text snippet method of fusion sequence grammer annotation framework |
CN110032729A (en) * | 2019-02-13 | 2019-07-19 | 北京航空航天大学 | A kind of autoabstract generation method based on neural Turing machine |
CN110598779A (en) * | 2017-11-30 | 2019-12-20 | 腾讯科技(深圳)有限公司 | Abstract description generation method and device, computer equipment and storage medium |
CN111192576A (en) * | 2018-11-14 | 2020-05-22 | 三星电子株式会社 | Decoding method, speech recognition device and system |
CN111386537A (en) * | 2017-10-27 | 2020-07-07 | 谷歌有限责任公司 | Decoder-only attention-based sequence-switched neural network |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105868829A (en) * | 2015-02-06 | 2016-08-17 | 谷歌公司 | Recurrent neural networks for data item generation |
CN105930314A (en) * | 2016-04-14 | 2016-09-07 | 清华大学 | Text summarization generation system and method based on coding-decoding deep neural networks |
-
2017
- 2017-02-23 CN CN201710099638.9A patent/CN106933785A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105868829A (en) * | 2015-02-06 | 2016-08-17 | 谷歌公司 | Recurrent neural networks for data item generation |
CN105930314A (en) * | 2016-04-14 | 2016-09-07 | 清华大学 | Text summarization generation system and method based on coding-decoding deep neural networks |
Non-Patent Citations (1)
Title |
---|
ALEX ALIFIMOFF: "Abstractive Sentence Summarization with Attentive Recurrent Neural Networks", 《PROCEEDINGS OF NAACL-HLT 2016》 * |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107766506A (en) * | 2017-10-20 | 2018-03-06 | 哈尔滨工业大学 | A kind of more wheel dialog model construction methods based on stratification notice mechanism |
CN111386537A (en) * | 2017-10-27 | 2020-07-07 | 谷歌有限责任公司 | Decoder-only attention-based sequence-switched neural network |
CN110598779B (en) * | 2017-11-30 | 2022-04-08 | 腾讯科技(深圳)有限公司 | Abstract description generation method and device, computer equipment and storage medium |
CN110598779A (en) * | 2017-11-30 | 2019-12-20 | 腾讯科技(深圳)有限公司 | Abstract description generation method and device, computer equipment and storage medium |
CN108153864A (en) * | 2017-12-25 | 2018-06-12 | 北京牡丹电子集团有限责任公司数字电视技术中心 | Method based on neural network generation text snippet |
CN108319668A (en) * | 2018-01-23 | 2018-07-24 | 义语智能科技(上海)有限公司 | Generate the method and apparatus of text snippet |
CN108319668B (en) * | 2018-01-23 | 2021-04-20 | 义语智能科技(上海)有限公司 | Method and equipment for generating text abstract |
CN108984524A (en) * | 2018-07-05 | 2018-12-11 | 北京理工大学 | A kind of title generation method based on variation neural network topic model |
CN109325110A (en) * | 2018-08-24 | 2019-02-12 | 广东外语外贸大学 | Indonesian documentation summary generation method, device, storage medium and terminal device |
CN109325110B (en) * | 2018-08-24 | 2021-06-25 | 广东外语外贸大学 | Indonesia document abstract generation method and device, storage medium and terminal equipment |
CN111192576A (en) * | 2018-11-14 | 2020-05-22 | 三星电子株式会社 | Decoding method, speech recognition device and system |
CN111192576B (en) * | 2018-11-14 | 2024-08-27 | 三星电子株式会社 | Decoding method, voice recognition device and system |
CN110032729A (en) * | 2019-02-13 | 2019-07-19 | 北京航空航天大学 | A kind of autoabstract generation method based on neural Turing machine |
CN109947930A (en) * | 2019-03-12 | 2019-06-28 | 上海秘塔网络科技有限公司 | Abstraction generating method, device, terminal and computer readable storage medium |
CN109948162A (en) * | 2019-03-25 | 2019-06-28 | 北京理工大学 | The production text snippet method of fusion sequence grammer annotation framework |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106933785A (en) | A kind of abstraction generating method based on recurrent neural network | |
CN106980683B (en) | Blog text abstract generating method based on deep learning | |
CN111897949B (en) | Guided text abstract generation method based on Transformer | |
CN107038159B (en) | A kind of neural network machine interpretation method based on unsupervised domain-adaptive | |
CN109635124B (en) | Remote supervision relation extraction method combined with background knowledge | |
CN106547735A (en) | The structure and using method of the dynamic word or word vector based on the context-aware of deep learning | |
CN108280112A (en) | Abstraction generating method, device and computer equipment | |
CN106383816B (en) | The recognition methods of Chinese minority area place name based on deep learning | |
CN109840322B (en) | Complete shape filling type reading understanding analysis model and method based on reinforcement learning | |
CN107133211A (en) | A kind of composition methods of marking based on notice mechanism | |
CN109635280A (en) | A kind of event extraction method based on mark | |
CN107273355A (en) | A kind of Chinese word vector generation method based on words joint training | |
CN107066973A (en) | A kind of video content description method of utilization spatio-temporal attention model | |
CN103325061B (en) | A kind of community discovery method and system | |
CN109271644A (en) | A kind of translation model training method and device | |
CN107133224A (en) | A kind of language generation method based on descriptor | |
CN107644014A (en) | A kind of name entity recognition method based on two-way LSTM and CRF | |
CN109977199B (en) | Reading understanding method based on attention pooling mechanism | |
CN108874997A (en) | A kind of name name entity recognition method towards film comment | |
CN109977234A (en) | A kind of knowledge mapping complementing method based on subject key words filtering | |
CN111897957B (en) | Capsule neural network integrating multi-scale feature attention and text classification method | |
CN107451278A (en) | Chinese Text Categorization based on more hidden layer extreme learning machines | |
Shah et al. | Image captioning using deep neural architectures | |
CN109145290A (en) | Based on word vector with from the semantic similarity calculation method of attention mechanism | |
CN107247703A (en) | Microblog emotional analysis method based on convolutional neural networks and integrated study |
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 | ||
AD01 | Patent right deemed abandoned |
Effective date of abandoning: 20201013 |
|
AD01 | Patent right deemed abandoned |