CN107766506A - A kind of more wheel dialog model construction methods based on stratification notice mechanism - Google Patents
A kind of more wheel dialog model construction methods based on stratification notice mechanism Download PDFInfo
- Publication number
- CN107766506A CN107766506A CN201710986813.6A CN201710986813A CN107766506A CN 107766506 A CN107766506 A CN 107766506A CN 201710986813 A CN201710986813 A CN 201710986813A CN 107766506 A CN107766506 A CN 107766506A
- Authority
- CN
- China
- Prior art keywords
- sentence
- word
- implicit function
- stratification
- encryption
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/332—Query formulation
- G06F16/3329—Natural language query formulation or dialogue systems
Abstract
The present invention relates to a kind of more wheel dialog model construction methods based on stratification notice mechanism, it is to depend on large-scale corpus to solve existing interactive system, training speed is influenceed by language material scale, and due to the reply nonuniqueness of dialogue generation, Seq2Seq models are invariably prone to generate general, the shortcomings that insignificant reply, and a kind of more wheel dialog model construction methods based on stratification notice mechanism are proposed, including：Sentence inputting is received, for each sentence, encryption implicit function is calculated since first word, calculates the Attention weights of each sentence, and calculates topic linguistic context and represents vector, finally calculates decryption implicit function, while result is exported.The present invention is applied to the chat robots system of open field.
Description
Technical field
The present invention relates to interactive system, and in particular to a kind of more wheel dialog models based on stratification notice mechanism
Construction method.
Background technology
1st, foreign technology present situation
(1) conversational system based on artificial template
Based on the technology of artificial template by manually setting session operational scenarios, and some are write to each scene and targetedly talked with
Template, template describe the problem of user is possible and corresponding answer template.
Weizenbaum et al. (1966) develops earliest chat robots ELIZA, ELIZA may go out according in dialogue
Existing language situation, goes to be pre-designed corresponding language template, and text generator can be according to the input of user by the weight in input
Want information to be embedded into template, finally give reply.
Chat is restricted to special scenes or specific topic by them, and generates sound using one group of pattern rule
Should.
(2) conversational system based on retrieval
Chat robots based on retrieval technique then using the method for being similar search engine, are previously stored dialogue storehouse simultaneously
Establish index, according to user's question sentence, in storehouse talk with progress fuzzy matching find most suitable response content.
Shaikh et al. (2010) constructs a cyberchat robot (VCA), can enter in chatroom with people
The preliminary social activity of row, they utilize ongoing conversation subject to carry out web search using a kind of novel method, and
Find the related subject that may be inserted into dialogue and change its flow, can be regarded as based on retrieval and and template method melt
Close.
(3) the dialogue generation model based on deep learning
Application of the depth learning technology in dialogue generates is mainly towards open field chat robots, because extensive logical
It is relatively easy to the acquisition of language material, the Sequence to Sequence models of the most frequently used reference machine translation, dialogue is given birth to
Into the whole process by problem to reply be considered as in machine translation translation process from original language to object language.
Ritter et al. (2011) has used the dialogue language material obtained from Twitter, to imitate using Seq2Seq models
Fruit has exceeded the dialog model based on searching system.
Sordoni et al. (2015) proposes a dialog generation system, and the system take into account the context letter in dialogue
Breath, so as to obtain lifting on uniformity is replied.
Serban et al. (2016) proposes Hierarchical Nerual Network models, it is intended in dialogue
Semanteme and interaction hereafter is modeled, so as to build the conversational system taken turns one more.
Jiwei Li et al. (2016) are directed to solving the problems, such as traditional Seq2Seq models generation common replies, introduce
Mutual information improves the diversity that generation is replied as object function.Jiwei Li (2016) use improved Seq2Seq simultaneously
Model is modeled to user styles, user embedding is introduced as priori in decoding end, so as to improve dialogue system
The uniformity and correlation of system.
Louis Shao et al. (2017) improve training method and the decode ends of Seq2Seq models, and add
Beam-search, so as to improve the reply length and uniformity and correlation of model generation.
2nd, domestic technique present situation
The country is because start late, the method that the research in terms of conversational system is also mainly based upon deep learning, Li
Hang et al. (2015) proposes Neural Responding Machine, using improved Seq2Seq models, adds
Attention mechanism simultaneously carries out fusion so as to achieve good result in short text conversational system using multiple models.
Mou Lili (2016), which are conceived to, solves the problems, such as traditional Seq2Seq models generation common replies, proposes first
Seq2Bf models, keyword is first predicted by using mutual information, then based on keyword reply the generation of sentence.
Zongcheng Ji (2014) then use the method based on retrieval simultaneously, using state-of-the-art information retrieval technique,
By huge dialogue corpus, a relatively intelligent conversational system is created.
3rd, the brief analysis of domestic and foreign literature summary
The research both at home and abroad for open field dialog generation system mainly includes the method based on template at present, based on retrieval
Method and method based on deep learning.The method based on template during early stage does not carry out real Language Processing, raw
Into language it is stiff, formalization, the problem of often existing in terms of semantic and fluency, relative to open field conversational system, the party
Method is more suitable for Task chat robots.
Method based on retrieval be it is existing everybody talk with corpus by sorting learning art and depth matches skill
Art finds the optimal reply for being adapted to currently input.The limitation of this method is to be only capable of being replied with fixed language mode, nothing
Method realizes the diversity combination of word.
Current most popular method is the method based on deep learning, uses the Seq2Seq moulds from machine translation task
Type, usually Encoder-Decoder structures, coordinate fairly large dialogue language material, trained end to end for realizing, so as to
Obtain a conversational system.This method can break through before limitation of the method for clause word, mainly user is inputted
The problem of be modeled, the generation of word for word (word) is then carried out according to intermediate result, can creatively generate replys, it is exhausted at present
Most of researchs are all based on the expansion or improvement of the model.
But the method based on deep learning depends on large-scale corpus, Seq2Seq model training speed is advised by language material
The influence of mould, and due to the reply nonuniqueness of dialogue generation, Seq2Seq models are invariably prone to generate general, it is not intended to justice
Reply, such as " hello ", " I does not also know ", " heartily " etc..
Furthermore current most of conversational systems are devoted to optimize the single-wheel dialogue i.e. quality of question-response process, and chat is
One continuous interactive process for having specific background, the meaning of a word will combine the background ability of the context of dialogue or correlation sometimes
It can determine that.The problem of needing to be studied is still for context modeling.
The content of the invention
The invention aims to solve existing interactive system to depend on large-scale corpus, training speed by
The influence of language material scale, and due to the reply nonuniqueness of dialogue generation, Seq2Seq models be invariably prone to generate it is general,
The shortcomings that insignificant reply, and a kind of more wheel dialog model construction methods based on stratification notice mechanism are proposed, including：
Step 1: receive n+1 sentence inputting co,c1,…cn；
Step 2: it is directed to each sentence ci, encryption implicit function h is calculated since first wordi,t=f (xi,t,
hi,t-1), wherein wherein xi,tRepresent ciT-th of word；Wherein hi,0It is designated as parameter preset；And finish last calculating
hi,tIt is designated as sentence ciEncryption implicit function hi；
Step 3: calculate the Attention weights of i-th of sentenceWherein ei=vTtanh(Whi+
Uhn)；V, W, U are the parameter preset in Attention mechanism；Tanh is activation primitive；
Step 4: calculating topic linguistic context represents vector T=∑ αihi；
Step 5: calculate decryption implicit function st=f (yt-1,st-1, T), yt-1Represent the iteration input quantity at t-1 moment, y0For
Preset value；s0=hn；
Step 6: by s1,s2,……st……snValue exported as a result.
The present invention also provides another more wheel dialog model construction methods based on stratification notice mechanism, including：
Step 1: receive n+1 sentence inputting co,c1,…cn；
Step 2: it is directed to each sentence ci, encryption implicit function h is calculated since first wordi,t=f (xi,t,
hi,t-1), wherein wherein xi,tRepresent ciT-th of word；Wherein hi,0It is designated as parameter preset；And finish last calculating
hi,tIt is designated as sentence ciEncryption implicit function hi；
Step 3: calculate the Attention weights of t-th of word in i-th of sentenceWherein eit=
vTtanh(Whi+Ust-1)；V, W, U are the parameter preset in Attention mechanism；st-1For the hidden layer state at t-1 moment；
Step 4: calculate dynamic representation vector Dt=αithi；
Step 5: calculate decryption implicit function st=f (yt-1,st-1, T), yt-1Represent the iteration input quantity at t-1 moment, y0For
Preset value；s0=hn；
Step 6: by s1,s2,……st……snValue exported as a result.
Beneficial effects of the present invention are：
1st, do not influenceed independent of large-scale corpus, training speed by language material scale, do not tend to generate it is general,
Insignificant reply；
2nd, the present invention is tested on Opensubtitles data sets and Ubuntu data sets.
On Opensubtitles data sets, Embedding Average of the invention can reach 0.565647,
Apparently higher than the 0.557139 of prior art；The Greedy Matching of the present invention can reach 0.523235, hence it is evident that be higher than
The 0.503273 of prior art；The Extrema of the present invention can reach 0.393724, higher than the 0.393189 of prior art.
On Ubuntu data sets, Embedding Average of the invention can reach 0.612089, hence it is evident that be higher than
The 0.577022 of prior art, Greedy Matching of the invention can reach 0.429328, hence it is evident that higher than prior art
0.416948, Extrema of the invention can reach 0.397543, higher than the 0.391392 of prior art.
Brief description of the drawings
Fig. 1 is the flow chart of more wheel dialog model construction methods based on stratification notice mechanism of the present invention；
Fig. 2 is the schematic diagram that encryption implicit function process is calculated in the step 2 of the specific embodiment of the invention one；Wherein
Context refers to sentence inputting co,c1,…cnThe context of composition；
Fig. 3 is schematic diagram of the step 3 of the specific embodiment of the invention one to step 5；Wherein Topic Net represent to calculate
Topic linguistic context represents the computing module of vector T, and topic vector represent that topic linguistic context represents vector；Decoder, which is used to calculate, to be solved
Close implicit function；By hnThe arrow for pointing to decoder represents that the calculating of decryption implicit function needs to use hnValue.
Embodiment
Embodiment one：
The present invention provides a kind of more wheel dialog model construction methods based on stratification notice mechanism, as shown in figure 1, bag
Include：
Step 1: receive n+1 sentence inputting co,c1,…cn。
Step 2: it is directed to each sentence ci, encryption implicit function h is calculated since first wordi,t=f (xi,t,
hi,t-1), wherein wherein xi,tRepresent ciT-th of word；Wherein hi,0It is designated as parameter preset；And finish last calculating
hi,tIt is designated as sentence ciEncryption implicit function hi。
Step 3: calculate the Attention weights of i-th of sentenceWherein ei=vTtanh(Whi+
Uhn)；V, W, U are the parameter preset in Attention mechanism.
Step 4: calculating topic linguistic context represents vector T=∑ αihi。
Step 5: calculate decryption implicit function st=f (yt-1,st-1, T), yt-1The iteration input quantity at t-1 moment is represented,0To be pre-
If value；s0=hn；When the sentence inputting received in step 1 is training data, yt-1For default model answer word, work as step
When the sentence inputting received in rapid one is measured data, yt-1Value be equal to st-1。
Step 6: by stValue exported as a result.
Specifically, mode of the invention is the Seq2Seq models based on machine translation task, and using
Encoder-Decoder structures, it is the characteristics of this model and structure, for a sentence ci, each list that will wherein include
Word xi,tModel is sequentially inputted, model is for each xi,tCorresponding encryption implicit function (as shown in Figure 2) is calculated according to formula, its
Result of calculation is used to calculate next encryption implicit function, and by that analogy, the value of last encryption implicit function output is as whole
The encryption implicit function h of sentencei, due to have input n+1 sentence altogether, therefore can also obtain n+1 encryption implicit function.Then root
Attention weights are calculated according to these encryption implicit functions, Attention mechanism is also conventional mechanism, and the present invention is to its formula
Form is modified, the improvement to formula is mainly when calculating the weight of each, it is contemplated that the encryption of current sentence
Implicit function (the i.e. formula e of last in implicit function and all sentencesi=vTtanh(Whi+Uhn) in hiRepresent current sentence
Encryption implicit function, hnRepresent last encryption implicit function), vector T then is represented further according to weight calculation topic linguistic context,
The vector contains the necessary information at encryption end, and this vector is decrypted last decrypting end (is used for the function i.e. s decryptedt
=f (yt-1,st-1, T)) result is obtained, as shown in Figure 3.During actual treatment data, yt-1Be output result, stIt is exactly yt, i.e. t
The word of moment output, and in the training process, in order to ensure the effect of training, stIt is actual output, and yt-1It is to preset
Word, i.e. word in model answer.
It can be seen that the method for the present invention is based not only on current problem and replied, current context and up and down also can take into account
Literary information, this process can be considered as to Memory Process, people read information from memory, with reference to current problem, so as to provide
Reply.This item work is handled whole context using memory network (Memory Network), remembers net end to end
Network can be used for QA tasks or read understanding, and different tasks is completed with this to obtain the expression of document by modeling.
Last sentence in context is considered as key (i.e. formula e by this item worki=vTtanh(Whi+Uhn) in hn), its
Remaining sentence is considered as memory, and representing for whole context is calculated with this, encryption end of the part as Seq2Seq models, as a result
Decrypting end is input to, is replied for decoding.The timing information of sentence in memory is considered that simultaneously, assigns each sentence
Different weights, represented with this to obtain more preferable context.
Embodiment two：
The present invention also provides another more wheel dialog model construction methods based on stratification notice mechanism, including：
Step 1: receive n+1 sentence inputting co,c1,…cn。
Step 2: it is directed to each sentence ci, encryption implicit function h is calculated since first wordi,t=f (xi,t,
hi,t-1), wherein wherein xi,tRepresent ciT-th of word；Wherein hi,0It is designated as parameter preset；And finish last calculating
hi,tIt is designated as sentence ciEncryption implicit function hi。
Step 3: calculate the Attention weights of t-th of word in i-th of sentenceWherein eit=
vTtanh(Whi+Ust-1)；V, W, U are the parameter preset in Attention mechanism；st-1For the hidden layer state at t-1 moment.
Step 4: calculate dynamic representation vector Dt=αithi。
Step 5: calculate decryption implicit function st=f (yt-1,st-1,Dt), yt-1Represent the iteration input quantity at t-1 moment, y0For
Preset value；s0=hn。
Step 6: by stValue exported as a result.
The difference of present embodiment and embodiment is, the side of Attention weights is calculated in step 3
Formula is different, and present embodiment is to calculate weight for each each word, and embodiment one be only for
Whole sentence calculates a weight.Another place's difference is that last encryption implicit function h is not used when weight specifically calculatesn, and
It is the hidden layer state s using last momentt-1.Embodiment one and embodiment two are all based on " considering on entire chapter
Hereafter, and from key values weight is calculated " this invention thinking, the key values only chosen one are to select last to add
Close implicit function, another is the hidden layer state for selecting decrypting end.
In the present invention, the method for deep learning has been used to be modeled contextual information, so as to optimize dialogue matter
Amount, improve correlation and uniformity that generation is replied.The main models that this item patent uses are Seq2Seq models, and structure is
Encoder-Decoder structures (encryption end-decryption end structure), because final goal is that decoding generation one is semantic smooth, one
The reply of cause property good relationship, Decoder (decrypting end) must be a preferable language model, so Decoder is
Realization based on RNN.
Traditional Seq2Seq models simply consider the problem of single-wheel dialogue and reply, but not only nationwide examination for graduation qualification when people engage in the dialogue
Dialogue is current, and what is said or talked about, also can take into account current context and contextual information, but in more wheel dialogues, current sentence is namely
Distance currently replys nearest sentence and is considered as most important sentence on time, because it is for the direct of the sentence that generation, which is replied,
Reply.In order to obtain the information of whole context, so as to be modeled to linguistic context topic, institute of this item work using RNN to context
There is sentence to be modeled, so as to obtain multiple expressions, then referring to the notice mechanism in machine translation, layer is carried out to context
The notice of secondaryization calculates, and show that the attented of context is represented, and in this, as the expression of topic linguistic context, be added to decoding
Hold and be used for auxiliary decoder, preferably replied so as to generate uniformity correlation.
Beneficial effects of the present invention are specifically described below：
The method main flow evaluated at present the chat robots under open field environment has two kinds of thinkings, objective indicator evaluation with
Subjective artificial scoring.The objective indicator part that this item work uses is mainly with Embedding Average, Greedy
Matching, Vector Extrema are the Evaluations matrix based on term vector of representative.
The general principle of above-mentioned objective indicator is all that the candidate for calculating generation is replied between the target retro of Key for Reference
Similarity, in this, as reply quality scoring, basic skills be by understand each word the meaning come judge reply
Correlation, term vector are the bases for realizing this evaluation method.According to semantic distribution, distribute a vector to each word and be used for
Represent this word, then by various methods can respectively obtain candidate reply with the sentence of target retro vector, then by cosine away from
From being compared, it is possible to obtain the similarity of the two.
Opensubtitles is the well-known captions website in a world, there is provided the captions of various countries' language are downloaded, and can therefrom be obtained
Obtain the English dialogue data collection being made up of film dialogue；Ubuntu data sets are that the chat record sticked is discussed in Ubuntu forums
The English dialogue data collection of composition.This patent work obtains 100000 Opensubtitles session operational scenarios, is processed into
The dialogue session of wheel more than 800000, on this basis using word2vec pre-training term vectors, has obtained 30000+ word
Table；The dialogue session of wheel more than about 450000 has then been obtained on Ubuntu data sets, has been obtained using word2vec pre-training term vectors
30000+ vocabulary is arrived.
At present topic model (method of embodiment one) and dynamic model (embodiment two
Method) tested on Opensubtitles data sets and Ubuntu data sets, by handling former dialogue language material
After context, dialogue session form is processed into, carries out experiment test.
This item patent work and three LSTM, HRED model, VHRED models baseline models are in Embedding
Compared in Average, Greedy Matching and Vector Extrema indexs.Test result indicates that this item is special
The topic model of profit work performances of three indexs on Opensubtitles data sets and Ubuntu data sets are better than
baseline；And dynamic model have also exceeded baseline performance in the first two index.It the results are shown in Table 1 and table 2.
Table 1:Result on Opensubtitles data sets, lstm, hred, vhred are baseline results, topic
Model, dynamic model are the method applied for a patent
Table 2:Result on Ubuntu data sets, lstm, hred, vhred are baseline results, topic model,
Dynamic model are the method applied for a patent
The effect of the present invention can be more intuitively found out by following 4 groups of test cases.
The method that wherein topic mode represent embodiment one, dyna model represent embodiment two
Method, hred model, vhred model and LSTM model are existing models.
Example 1：
Above:
I will attempt the exploitation version using PPA now, look at whether it can collapse again.
You are seeing the cpu temperature of computer
No, the problem of I is not run into also in terms of temperature ... I where can see its temperature
topic model：You can attempt to delete your configuration file, then try again
dyna model：Attempt from terminal operating lspci, there is individual list there
hred model：System->Preference->Power management
vhred model：I does not understand your meaning
LSTM model：I does not understand
Example 2：
Above：
Come in.
It thanks.
It is not so polite, it is in own home just to work as
Miss Di La etc. be able to can go out.
Alright
Have fun
topic model:Had fun I believe that can do by myself
dyna model:Thank you
hred model:Come on
vhred model:Alright, we walk
lstm model:I understands careful
Example 3：
Above:
According to the method in link, nomodeset parameters are added in kernel into startup item
In additionIt thanks.
topic model：You can also use nomodeset options editor's configuration text under/etc/default/grub
Part
dyna model：Instruct identical with last entry of last time renewal
hred model：It is unfriendly
vhred model：It is unfriendly
LSTM model：I does not understand
Example 4：
Above：
Who knows another problem.Pidgin is allowed to show messages menu either with or without method
It please check down whether you really use NVIDIA driver.Or reset driving ensure to have no problem
topic model：I has a nvidia video card, and a video card of Radeon Mobility 7000.
dyna model：I has one 75 feet of best display.
hred model：How I this do
vhred model：I knows
lstm model：I does not know
Vhred model and lstm model are can be seen that from above-mentioned 4 examples to be more likely to generate general reply,
It is difficult to the context generation for complexity meets the reply of linguistic context.Hred model generate general reply sometimes, sometimes
Generation meets the reply of linguistic context, but generally speaking, the situation for not meeting linguistic context is more.And the present invention topic model and
Dyna model can generate the reply content for meeting linguistic context.
The embodiment one, two of the present invention may be directly applied in the chat robots system of open field.This is specially
The chat robots that the application vector of sharp technology is Harbin Institute of Technology's social computing with Research into information retrieval center is developed are " stupid
It is stupid ".
More wheel dialogue functions provided by the invention can be as a module of whole chat robots system：The module connects
By the enabling signal from middle control module, independently it is responsible for once the progress of more wheel dialogues, when taking turns end-of-dialogue, by control flow more
Give back middle control module.
From deployment way, this technology can independently be used as a calculate node, be deployed in Ali's cloud or U.S.
On the cloud computing platforms such as group's cloud, the communication between other modules can be carried out by way of binding IP address and port numbers.
In the specific implementation of the technology, because having used deep learning correlation technique, corresponding depth is needed to use
Learning framework：The all alternatively frameworks of Tensorflow and Pytorch.No matter any deep learning framework is used, not
The external interface of the technology modules can be influenceed.
The present invention can also have other various embodiments, in the case of without departing substantially from spirit of the invention and its essence, this area
Technical staff works as can make various corresponding changes and deformation according to the present invention, but these corresponding changes and deformation should all belong to
The protection domain of appended claims of the invention.
Claims (4)
1. a kind of more wheel dialog model construction methods based on stratification notice mechanism, including：
Step 1: receive n+1 sentence inputting co, c1... cn；
Step 2: it is directed to each sentence ci, encryption implicit function h is calculated since first wordI, t=f (xI, t, hI, t-1), its
In wherein xI, tRepresent ciT-th of word；Wherein hI, 0For parameter preset；And the h for finishing last calculatingI, tAs sentence ci
Encryption implicit function hi；
Step 3: calculate the Attention weights of i-th of sentenceWherein ei=vTtanh(Whi+Uhn)；v、
W, U is the parameter preset in Attention mechanism；Tanh is activation primitive；
Step 4: calculating topic linguistic context represents vector T=∑ αihi；
Step 5: calculate decryption implicit function st=f (yt-1, st-1, T), yt-1Represent the iteration input quantity at t-1 moment, y0It is default
Value；s0=hn；
Step 6: by s1, s2... snValue exported as a result.
2. more wheel dialog model construction methods according to claim 1 based on stratification notice mechanism, its feature exist
In, when the sentence inputting received in step 1 is training data, y in step 5t-1For default model answer word, work as step
When the sentence inputting received in rapid one is measured data, y in step 5t-1Value be equal to st-1。
3. a kind of more wheel dialog model construction methods based on stratification notice mechanism, including：
Step 1: receive n+1 sentence inputting co, c1... cn；
Step 2: it is directed to each sentence ci, encryption implicit function h is calculated since first wordI, t=f (xI, t, hI, t-1), its
In wherein xI, tRepresent ciT-th of word；Wherein hI, 0For parameter preset；And the h for finishing last calculatingI, tAs sentence ci
Encryption implicit function hi；
Step 3: calculate the Attention weights of t-th of word in i-th of sentenceWherein eit=
vTtanh(Whi+Ust-1)；V, W, U are the parameter preset in Attention mechanism；st-1For the hidden layer state at t-1 moment；
Tanh is activation primitive；
Step 4: calculate dynamic representation vector Dt=αithi；
Step 5: calculate decryption implicit function st=f (yt-1, st-1, Dt), yt-1Represent the iteration input quantity at t-1 moment, y0It is default
Value；s0=hn；
Step 6: by s1, s2... snValue exported as a result.
4. more wheel dialog model construction methods according to claim 3 based on stratification notice mechanism, its feature exist
In, when the sentence inputting received in step 1 is training data, y in step 5t-1For default model answer word, work as step
When the sentence inputting received in rapid one is measured data, y in step 5t-1Value be equal to st-1。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710986813.6A CN107766506A (en) | 2017-10-20 | 2017-10-20 | A kind of more wheel dialog model construction methods based on stratification notice mechanism |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710986813.6A CN107766506A (en) | 2017-10-20 | 2017-10-20 | A kind of more wheel dialog model construction methods based on stratification notice mechanism |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107766506A true CN107766506A (en) | 2018-03-06 |
Family
ID=61268560
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710986813.6A Pending CN107766506A (en) | 2017-10-20 | 2017-10-20 | A kind of more wheel dialog model construction methods based on stratification notice mechanism |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107766506A (en) |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108388944A (en) * | 2017-11-30 | 2018-08-10 | 中国科学院计算技术研究所 | LSTM neural network chips and its application method |
| CN108491514A (en) * | 2018-03-26 | 2018-09-04 | 清华大学 | The method and device putd question in conversational system, electronic equipment, computer-readable medium |
| CN108563628A (en) * | 2018-03-07 | 2018-09-21 | 中山大学 | Talk with generation method based on the emotion of HRED and inside and outside memory network unit |
| CN108805088A (en) * | 2018-06-14 | 2018-11-13 | 南京云思创智信息科技有限公司 | Physiological signal analyzing subsystem based on multi-modal Emotion identification system |
| CN108920510A (en) * | 2018-05-30 | 2018-11-30 | 出门问问信息科技有限公司 | Automatic chatting method, device and electronic equipment |
| CN108959246A (en) * | 2018-06-12 | 2018-12-07 | 北京慧闻科技发展有限公司 | Answer selection method, device and electronic equipment based on improved attention mechanism |
| CN109658270A (en) * | 2018-12-19 | 2019-04-19 | 前海企保科技（深圳）有限公司 | It is a kind of to read the core compensation system and method understood based on insurance products |
| CN109684912A (en) * | 2018-11-09 | 2019-04-26 | 中国科学院计算技术研究所 | A kind of video presentation method and system based on information loss function |
| CN109948089A (en) * | 2019-02-21 | 2019-06-28 | 中国海洋大学 | A kind of method and device for extracting Web page text |
| CN110297894A (en) * | 2019-05-22 | 2019-10-01 | 同济大学 | A kind of Intelligent dialogue generation method based on auxiliary network |
| CN110334190A (en) * | 2019-07-12 | 2019-10-15 | 电子科技大学 | A kind of reply automatic generation method towards open field conversational system |
| CN110347833A (en) * | 2019-07-09 | 2019-10-18 | 浙江工业大学 | A kind of classification method of more wheel dialogues |
| CN110413729A (en) * | 2019-06-25 | 2019-11-05 | 江南大学 | Talk with generation method based on the more wheels of tail sentence-dual attention model of context |
| CN110516213A (en) * | 2019-09-03 | 2019-11-29 | 哈尔滨工业大学 | A kind of structural data generation text method based on table hierarchyization modeling |
| CN111797218A (en) * | 2020-07-07 | 2020-10-20 | 海南中智信信息技术有限公司 | Open domain dialogue generation method based on Cycle-Seq2Seq |
| CN111914983A (en) * | 2019-05-07 | 2020-11-10 | 阿里巴巴集团控股有限公司 | Interaction method and device, sound box, electronic equipment and storage medium |
| CN112417089A (en) * | 2019-08-21 | 2021-02-26 | 东北大学秦皇岛分校 | High-parallelism reading understanding method based on deep learning |
| CN113488036A (en) * | 2020-06-10 | 2021-10-08 | 海信集团有限公司 | Multi-round voice interaction method, terminal and server |
| US11245648B1 (en) | 2020-07-31 | 2022-02-08 | International Business Machines Corporation | Cognitive management of context switching for multiple-round dialogues |
| CN114357129A (en) * | 2021-12-07 | 2022-04-15 | 华南理工大学 | High-concurrency multi-round chat robot system and data processing method thereof |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106126596A (en) * | 2016-06-20 | 2016-11-16 | 中国科学院自动化研究所 | A kind of answering method based on stratification memory network |
| CN106354710A (en) * | 2016-08-18 | 2017-01-25 | 清华大学 | Neural network relation extracting method |
| CN106383815A (en) * | 2016-09-20 | 2017-02-08 | 清华大学 | Neural network sentiment analysis method in combination with user and product information |
| CN106776578A (en) * | 2017-01-03 | 2017-05-31 | 竹间智能科技（上海）有限公司 | Talk with the method and device of performance for lifting conversational system |
| CN106844368A (en) * | 2015-12-03 | 2017-06-13 | 华为技术有限公司 | For interactive method, nerve network system and user equipment |
| CN106847271A (en) * | 2016-12-12 | 2017-06-13 | 北京光年无限科技有限公司 | A kind of data processing method and device for talking with interactive system |
| CN106933785A (en) * | 2017-02-23 | 2017-07-07 | 中山大学 | A kind of abstraction generating method based on recurrent neural network |
| CN107133211A (en) * | 2017-04-26 | 2017-09-05 | 中国人民大学 | A kind of composition methods of marking based on notice mechanism |
| CN107256228A (en) * | 2017-05-02 | 2017-10-17 | 清华大学 | Answer selection system and method based on structuring notice mechanism |
-
2017
- 2017-10-20 CN CN201710986813.6A patent/CN107766506A/en active Pending
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106844368A (en) * | 2015-12-03 | 2017-06-13 | 华为技术有限公司 | For interactive method, nerve network system and user equipment |
| CN106126596A (en) * | 2016-06-20 | 2016-11-16 | 中国科学院自动化研究所 | A kind of answering method based on stratification memory network |
| CN106354710A (en) * | 2016-08-18 | 2017-01-25 | 清华大学 | Neural network relation extracting method |
| CN106383815A (en) * | 2016-09-20 | 2017-02-08 | 清华大学 | Neural network sentiment analysis method in combination with user and product information |
| CN106847271A (en) * | 2016-12-12 | 2017-06-13 | 北京光年无限科技有限公司 | A kind of data processing method and device for talking with interactive system |
| CN106776578A (en) * | 2017-01-03 | 2017-05-31 | 竹间智能科技（上海）有限公司 | Talk with the method and device of performance for lifting conversational system |
| CN106933785A (en) * | 2017-02-23 | 2017-07-07 | 中山大学 | A kind of abstraction generating method based on recurrent neural network |
| CN107133211A (en) * | 2017-04-26 | 2017-09-05 | 中国人民大学 | A kind of composition methods of marking based on notice mechanism |
| CN107256228A (en) * | 2017-05-02 | 2017-10-17 | 清华大学 | Answer selection system and method based on structuring notice mechanism |
Non-Patent Citations (4)
| Title |
|---|
| DZMITRY BAHDANAU: "Neural Machine Translation by Jointly Learning to Align and Translate", 《COMPUTER SCIENCE》 * |
| HUAYU LI等: "A Context-aware Attention Network for Interactive Question Answering", 《KDD "17: PROCEEDINGS OF THE 23RD ACM SIGKDD INTERNATIONAL CONFERENCE ON KNOWLEDGE DISCOVERY AND DATA MINING》 * |
| LIFENG SHANG: "Neural Responding Machine for Short-Text Conversation", 《ACL》 * |
| ZICHAO YANG等: "Hierarchical Attention Networks for Document Classification", 《PROCEEDINGS OF THE 2016 CONFERENCE OF THE NORTH AMERICAN CHAPTER OF THE ASSOCIATION FOR COMPUTATIONAL LINGUISTICS: HUMAN LANGUAGE TECHNOLOGIES》 * |
Cited By (29)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108388944A (en) * | 2017-11-30 | 2018-08-10 | 中国科学院计算技术研究所 | LSTM neural network chips and its application method |
| CN108563628A (en) * | 2018-03-07 | 2018-09-21 | 中山大学 | Talk with generation method based on the emotion of HRED and inside and outside memory network unit |
| CN108491514A (en) * | 2018-03-26 | 2018-09-04 | 清华大学 | The method and device putd question in conversational system, electronic equipment, computer-readable medium |
| CN108920510A (en) * | 2018-05-30 | 2018-11-30 | 出门问问信息科技有限公司 | Automatic chatting method, device and electronic equipment |
| CN108959246A (en) * | 2018-06-12 | 2018-12-07 | 北京慧闻科技发展有限公司 | Answer selection method, device and electronic equipment based on improved attention mechanism |
| CN108959246B (en) * | 2018-06-12 | 2022-07-12 | 北京慧闻科技(集团)有限公司 | Answer selection method and device based on improved attention mechanism and electronic equipment |
| CN108805088A (en) * | 2018-06-14 | 2018-11-13 | 南京云思创智信息科技有限公司 | Physiological signal analyzing subsystem based on multi-modal Emotion identification system |
| CN108805088B (en) * | 2018-06-14 | 2021-05-28 | 南京云思创智信息科技有限公司 | Physiological signal analysis subsystem based on multi-modal emotion recognition system |
| CN109684912A (en) * | 2018-11-09 | 2019-04-26 | 中国科学院计算技术研究所 | A kind of video presentation method and system based on information loss function |
| CN109658270A (en) * | 2018-12-19 | 2019-04-19 | 前海企保科技（深圳）有限公司 | It is a kind of to read the core compensation system and method understood based on insurance products |
| CN109948089A (en) * | 2019-02-21 | 2019-06-28 | 中国海洋大学 | A kind of method and device for extracting Web page text |
| CN111914983A (en) * | 2019-05-07 | 2020-11-10 | 阿里巴巴集团控股有限公司 | Interaction method and device, sound box, electronic equipment and storage medium |
| CN111914983B (en) * | 2019-05-07 | 2023-10-24 | 阿里巴巴集团控股有限公司 | Interaction method and device, sound box, electronic equipment and storage medium |
| CN110297894A (en) * | 2019-05-22 | 2019-10-01 | 同济大学 | A kind of Intelligent dialogue generation method based on auxiliary network |
| CN110297894B (en) * | 2019-05-22 | 2021-03-26 | 同济大学 | Intelligent dialogue generating method based on auxiliary network |
| CN110413729B (en) * | 2019-06-25 | 2023-04-07 | 江南大学 | Multi-turn dialogue generation method based on clause-context dual attention model |
| CN110413729A (en) * | 2019-06-25 | 2019-11-05 | 江南大学 | Talk with generation method based on the more wheels of tail sentence-dual attention model of context |
| CN110347833A (en) * | 2019-07-09 | 2019-10-18 | 浙江工业大学 | A kind of classification method of more wheel dialogues |
| CN110334190A (en) * | 2019-07-12 | 2019-10-15 | 电子科技大学 | A kind of reply automatic generation method towards open field conversational system |
| CN112417089A (en) * | 2019-08-21 | 2021-02-26 | 东北大学秦皇岛分校 | High-parallelism reading understanding method based on deep learning |
| CN112417089B (en) * | 2019-08-21 | 2022-12-09 | 东北大学秦皇岛分校 | High-parallelism reading understanding method based on deep learning |
| CN110516213A (en) * | 2019-09-03 | 2019-11-29 | 哈尔滨工业大学 | A kind of structural data generation text method based on table hierarchyization modeling |
| CN110516213B (en) * | 2019-09-03 | 2022-04-15 | 哈尔滨工业大学 | Structured data text generation method based on table hierarchical modeling |
| CN113488036A (en) * | 2020-06-10 | 2021-10-08 | 海信集团有限公司 | Multi-round voice interaction method, terminal and server |
| CN111797218A (en) * | 2020-07-07 | 2020-10-20 | 海南中智信信息技术有限公司 | Open domain dialogue generation method based on Cycle-Seq2Seq |
| CN111797218B (en) * | 2020-07-07 | 2022-03-29 | 海南中智信信息技术有限公司 | Open domain dialogue generation method based on Cycle-Seq2Seq |
| US11245648B1 (en) | 2020-07-31 | 2022-02-08 | International Business Machines Corporation | Cognitive management of context switching for multiple-round dialogues |
| CN114357129A (en) * | 2021-12-07 | 2022-04-15 | 华南理工大学 | High-concurrency multi-round chat robot system and data processing method thereof |
| CN114357129B (en) * | 2021-12-07 | 2023-02-14 | 华南理工大学 | High-concurrency multi-round chat robot system and data processing method thereof |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107766506A (en) | A kind of more wheel dialog model construction methods based on stratification notice mechanism | |
| Nan et al. | Improving factual consistency of abstractive summarization via question answering | |
| CN109460463B (en) | Model training method, device, terminal and storage medium based on data processing | |
| CN109271493A (en) | A kind of language text processing method, device and storage medium | |
| CN109547331A (en) | One kind is based on more wheel voice-enabled chat model building methods | |
| CN107423274A (en) | Commentary content generating method, device and storage medium based on artificial intelligence | |
| CN109791549A (en) | Machine customer interaction towards dialogue | |
| CN109964223A (en) | Session information processing method and its device, storage medium | |
| CN110234018B (en) | Multimedia content description generation method, training method, device, equipment and medium | |
| CN109284502B (en) | Text similarity calculation method and device, electronic equipment and storage medium | |
| CN110427461A (en) | Intelligent answer information processing method, electronic equipment and computer readable storage medium | |
| CN107679225B (en) | Reply generation method based on keywords | |
| CN111672098A (en) | Virtual object marking method and device, electronic equipment and storage medium | |
| CN108960574A (en) | Quality determination method, device, server and the storage medium of question and answer | |
| CN108345612A (en) | A kind of question processing method and device, a kind of device for issue handling | |
| CN110457661A (en) | Spatial term method, apparatus, equipment and storage medium | |
| CN113761156A (en) | Data processing method, device and medium for man-machine interaction conversation and electronic equipment | |
| CN110427454A (en) | Text mood analysis method and device, electronic equipment and non-transient storage media | |
| CN116956116A (en) | Text processing method and device, storage medium and electronic equipment | |
| Jhan et al. | CheerBots: Chatbots toward empathy and emotionusing reinforcement learning | |
| CN107729983A (en) | A kind of method, apparatus and electronic equipment using realizing of Robot Vision man-machine chess | |
| CN116258147A (en) | Multimode comment emotion analysis method and system based on heterogram convolution | |
| CN110491372A (en) | A kind of feedback information generating method, device, storage medium and smart machine | |
| CN116975214A (en) | Text generation method, device, storage medium and computer equipment | |
| CN113392640B (en) | Title determination method, device, equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |