CN109710946A - A kind of joint debate digging system and method based on dependence analytic tree - Google Patents
A kind of joint debate digging system and method based on dependence analytic tree Download PDFInfo
- Publication number
- CN109710946A CN109710946A CN201910034772.XA CN201910034772A CN109710946A CN 109710946 A CN109710946 A CN 109710946A CN 201910034772 A CN201910034772 A CN 201910034772A CN 109710946 A CN109710946 A CN 109710946A
- Authority
- CN
- China
- Prior art keywords
- argument
- text
- vector
- debate
- word
- 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
Abstract
The present invention relates to a kind of based on the joint debate digging system for relying on analytic tree, comprising: data preprocessing module, for being pre-processed to data;Text is embedded in module, for from extracting in the text of input, word, character, part of speech, the vector of dependence and argument type is indicated between argument;Sequential coding module, using the contextual information of two-way length Memory Neural Networks learning text in short-term, for completing the task of argument border detection and argument Relation extraction;Analytic tree module is relied on, analytic tree is relied on by building, for finding shortest path warp in two argument component entities;Label output module is excavated in debate, and the Tag Estimation work of three tasks, the type label of argument and the relational tags of argument are excavated for completing debate.The present invention can practise the text vector feature of high quality from debate text data middle school, finally detect the debate structure of text.
Description
Technical field
The present invention relates to natural language processing fields, and in particular to a kind of to excavate system based on the joint debate for relying on analytic tree
System and method.
Background technique
Currently, many technical methods can be used for debate excavation.Traditional debate method for digging is all mainly to subtask
Independent modeling, and the related information between three subtasks is had ignored, lead to degraded performance.In addition, there are also part work to use
Pipeline model carries out joint modeling to three subtasks, these models have error propagation in the training process.
Currently, there is some research methods based on assembly line.Its basic idea is three sons excavated to debate
The method that task uses assembly line is solved according to the sequence of assembly line.Wrong meeting of the pipelining technique due to argument type identification
, there is error propagation in the extraction mistake for influencing argument relationship.In addition, this method will identify that the argument come carries out
It matches two-by-two, carries out the classification of argument relationship later, produce the redundancy of argument relationship pair.
However, debate Research on Mining method often has ignored the related information between subtask at present, it is every there is also ignoring
The problem of a subtask different characteristics, and related information has critically important meaning to debate excavation, the label of a task is pre-
Surveying result can be used as the validity feature for predicting that subtask label is excavated in other debates.Therefore, in view of the above deficiencies, it is desirable to
It is a kind of more efficient, careful and can make full use of between subtask related information and every height is made full use of to appoint to find
The method of business feature, and then improve the performance of debate mining model.
Summary of the invention
In view of this, the purpose of the present invention is to provide a kind of based on the joint debate digging system for relying on analytic tree and side
Method can practise the text vector feature of high quality from debate text data middle school, finally detect the debate structure of text.
To achieve the above object, the present invention adopts the following technical scheme:
A kind of joint debate digging system based on dependence analytic tree, comprising:
One data preprocessing module, for being pre-processed to data;
One text is embedded in module, for from extracting word, character, part of speech, dependence between argument in the text of input
And the vector of argument type indicates;
One sequential coding module has been used to using the contextual information of two-way length Memory Neural Networks learning text in short-term
At the task of argument border detection and argument Relation extraction;
One relies on analytic tree module, relies on analytic tree by building, most short for finding in two argument component entities
Lu Jing;
Label output module is excavated in one debate, and the Tag Estimation work of three tasks is excavated for completing debate, argument
The relational tags of type label and argument.
Further, the data preprocessing module pre-process to data and be specifically included:
(1) web page interlinkage in document, spcial character, punctuation mark are removed;
(2) word segmentation processing is carried out to document;
(3) stem reduction treatment is carried out to English data;
(4) stop words for including in data set is filtered out according to the deactivated vocabulary of Chinese and English respectively.
Further, the text insertion module uses depth convolutional neural networks.
Further, a kind of analytic method based on the joint debate digging system for relying on analytic tree, feature exist
In, comprising the following steps:
Step S1: being input to data preprocessing module for opinion distortion document to be excavated and pre-process, obtained pretreatment
Text afterwards, and input text insertion module;
Step S2: text is embedded in module using depth convolutional neural networks to pretreated Text Feature Extraction word, character, word
Property, the vector of dependence and argument type indicates between argument, and list entries coding module;
Step S3: list entries coding module is embedded in the text data that module inputs according to text, in short-term using two-way length
The contextual information of Memory Neural Networks learning text data completes argument border detection and argument Relation extraction, obtains argument
The type label of component entity;
Step S4: dependence is closed according to relying between the type label and argument of argument component entity, relies on analytic tree
Module relies on analytic tree by building, and training obtains the label of argument relationship;
Step S5: label output module is excavated in debate, and the relational tags of the type label of obtained argument and argument are defeated
Out.
Further, the step S2 specifically:
Step S21: the input of depth convolutional neural networks is pretreated text sequence x=[x1,x2,...,xn], according to
The sequence of word in text sentence, every a line are all the words indicated by d dimensional vector, and CNN output is sequence C=[c1,
c2,...,cn,], C indicates to input the feature of each word, and n indicates the maximum length of list entries;
Step S22: the convolution kernel W ∈ R for the use of one width of narrow convolution sum being k between x(d×k), and willWithThe head and tail portion of sequence are filled into as filling vector;
Step S23: Text Feature Extraction word V is exported respectivelyw, character Vc, part of speech Vp, dependence V between argumentdAnd argument
Type VeVector indicate, and in list entries coding module.
Further, the step S3 specifically:
Step S31: the input of sequential coding layer is the sharing feature parameter vector of text embeding layer output, including word Vw, word
Accord with Vc, part of speech Vp, for learning text contextual information and identify argument component entity;
Step S32: one two-way LSTM of building calculates and obtains sentence vector, each LSTM unit is in t-th of word by one
A n-dimensional vector composition, comprising: an input gate it, a forgetting door ft, an out gate ot, a memory unit ct, and
One hidden unit ht, the vector input of one n dimension of each LSTM unit reception, previous hidden state is ht-1, previous note
Recalling unit is ct-1;
Undated parameter according to the following formula:
it=σ (W(i)xt+I(i)ht-1+bi)
ft=σ (W(f)xt+I(f)ht-1+bf)
ot=σ (W(o)xt+I(o)ht-1+bo)
ut=tanh (W(u)xt+I(u)ht-1+bu)
ct=it⊙ut+ft⊙ct-1
ht=ot⊙tanh(ct)
Wherein, σ indicates that logistic activation primitive, ⊙ indicate that the dot product of vector, W and I indicate that weight matrix, b indicate inclined
Difference vector, input of the LSTM unit on t-th of word are the word V of t-th of wordt w, character Vt cWith part of speech Vt pConnection vectorBy the hidden unit of two reversed LSTMWithIt is connected asAs output;
Step S32: upper BIO label is marked to each word of input sentence, its argument type is then marked again, is formed
The form of " BIO- argument type ";
Step S33: one two layers of neural network being made of DenseNet and Softmax of building:
Wherein, W is weight matrix, and b is bias vector;
Step S34: by stWith the vector e of previous wordi-1As input, it is input to later by one layer of neural network
The type label of Softmax layers of acquisition argument component entity obtains output and is mapped as vector ei。
Further, the step S4 specifically:
Step S41: by the type label e of argument component entityiWith the dependence V from text embeding layereInput dependence
Analytic sheaf;
Step S42: the LSTM combination recurrent neural network of two-way tree construction is constructed, and by following formula in LSTM unit
The interior n-dimensional vector for calculating t-th of node:
ht=ot⊙tanh(ct)
Wherein, m () is mapping function, and C (t) is the child node of t-th of node, and i is shared parameter.
The relationship between two target words pair is indicated using shortest path structure, it is for capturing between target word pair
Independent path will rely on analytic sheaf and be stacked on sequence layer, text sequence and the information for relying on analytic tree are merged into output
In, rely on the LSTM input of t-th of word of analytic sheaf are as follows:Hidden unit s in catenation sequence layertAnd opinion
Point relational dependence typeAnd the entity of argument component indicates
The direction of the type of relationship and relationship: being indicated the relationship between argument by step S43, and the dependence of each candidate is closed
System can be expressed as dp=[↑ hpA:↓hp1:↓hp2], wherein ↓ hpAIndicate the corresponding minimum father node of two argument entity nodes
Hidden layer, ↑ hp1With ↑ hp2It is the hidden state vector of two LSTM units, respectively indicates first in top-down LSTM-RNN
With second target argument physical components;
Step S44: one two layers of neural network of setting, it includes the hidden layer h of n dimension(r)With Softmax's
Output layer:
Wherein, W is weight matrix, and b is bias vector;
The LSTM-RNNs of tree construction is superimposed upon on sequence layer, the input d of argument relationship classification is constructedp, will be each
The average value of the hidden state vector of argument physical components is connected to d from sequence layerpCarry out the classification of argument relationship, obtain as
Lower formula:
Wherein, Up1And Up2It is the index of first and second argument entity set of words;
Step S45: being to be assigned with two labels to according to direction to each word in prediction, when the both direction mark of prediction
When signing inconsistent, select the relationship with high confidence as the final result of output, and training output obtains argument relationship
Label.
Further, the two-way mode referred to from top to bottom and from the bottom up of the two-way tree construction, not only to each
Node transmits the information from leaf node, also transmits information to node
Compared with the prior art, the invention has the following beneficial effects:
The automatic system that must be identified argument and extract the relationship between argument in subjectivity document of the present invention can combine more
Tasking learning method carries out debate excavation from debate text in high quality.
Detailed description of the invention
Fig. 1 is the schematic configuration view of present system.
Specific embodiment
The present invention will be further described with reference to the accompanying drawings and embodiments.
Fig. 1 is please referred to, the present invention provides a kind of based on the joint debate digging system for relying on analytic tree, comprising:
One data preprocessing module, for being pre-processed to data;
One text is embedded in module, for from extracting word, character, part of speech, dependence between argument in the text of input
And the vector of argument type indicates;
One sequential coding module has been used to using the contextual information of two-way length Memory Neural Networks learning text in short-term
At the task of argument border detection and argument Relation extraction;
One relies on analytic tree module, relies on analytic tree by building, most short for finding in two argument component entities
Lu Jing;
Label output module is excavated in one debate, and the Tag Estimation work of three tasks is excavated for completing debate, argument
The relational tags of type label and argument.
In this implementation, each module concrete function are as follows:
(1) data preprocessing module
Information abundant is contained in the online debate document of input but while being also mingled with certain noise.Cause
This, first pre-processes data, is substantially carried out the operation of the following aspects:
1. removing the web page interlinkage in document, spcial character, punctuation mark etc.;
2. pair document carries out word segmentation processing;
3. a pair English data carry out stem reduction treatment;
4. filtering out the stop words for including in data set respectively according to the deactivated vocabulary of Chinese and English.
2) text is embedded in module
Extracted from the text of input using convolutional neural networks word, character, part of speech (Part-of-Speech), argument it
Between dependence and argument type expression, depth convolutional neural networks (CNN) input be text sequence x=[x1,
x2,...,xn], according to the sequence of word in text sentence, every a line is all the word indicated by d dimensional vector, CNN output
For sequence C=[c1,c2,...,cn,], C indicates to input the feature of each word, and n indicates the maximum length of list entries.We
The convolution kernel W ∈ R for the use of one width of narrow convolution sum being k between x(d×k), and willWithAs filling
Vector is filled into the head and tail portion of sequence, to guarantee that the length of list entries will not change after convolutional layer.Point
It Shu Chu not Vw, Vc, Vp, VdAnd Ve, these parameters input in subsequent text sequence layer as the bottom shared parameter of model and carry out
Training study.
3) sequence layer module
The input of sequential coding layer is the sharing feature parameter vector of text embeding layer output, carrys out the context letter of learning text
It ceases and identifies argument component entity, calculated using a two-way LSTM obtain sentence vector first, each LSTM unit is at t-th
Word is made of a n-dimensional vector, comprising: input gate (input gate) it, forgetting door (forget gate) ft,
One out gate (output gate) ot, memory unit (memory cell) ctAn and hidden unit ht, each
LSTM unit receives the vector input of a n dimension, and previous hidden state is ht-1, previous memory unit is ct-1.According to
Lower formula undated parameter:
it=σ (W(i)xt+I(i)ht-1+bi)
ft=σ (W(f)xt+I(f)ht-1+bf)
ot=σ (W(o)xt+I(o)ht-1+bo)
ut=tanh (W(u)xt+I(u)ht-1+bu)
ct=it⊙ut+ft⊙ct-1
ht=ot⊙tanh(ct)
Wherein, σ indicates that logistic activation primitive, ⊙ indicate that the dot product of vector, W and I indicate that weight matrix, b indicate inclined
Difference vector, input of the LSTM unit on t-th of word are the word V of t-th of wordt w, character Vt cWith part of speech Vt pConnection vectorWe are simultaneously by the hidden unit of two reversed LSTMWithIt is connected asAs output.
All regard the identification for the two argument type of one argument border detection of task and task that debate is excavated as sequence labelling to ask
Topic, we first mark upper BIO label to each word of input sentence, then mark its argument type, i.e., each word again
The form of " BIO- argument type " is formed, such labeling method is both the label of debate mining task one and task two.In sequence
The top layer of column coding layer completes this two tasks, we realize two layers of nerve being made of DenseNet and Softmax
Network:
Wherein, W is weight matrix, and b is bias vector.
In the decoding process of argument Entity recognition, it is contemplated that the dependence of label is come pre- using the predicted value of a word
Survey the value of next word, specific practice is us by stWith the vector e of previous wordi-1As input, pass through one layer of mind later
The type label for obtaining argument component entity to Softmax layers through network inputs obtains output and is mapped as vector ei。
4) analytic sheaf module is relied on
Rely on the type mark that the input of analytic sheaf module is the argument component entity that sequence layer neural metwork training exports
Sign eiWith the dependence V from text embeding layere。
It is realized using the mode of the LSTM combination recurrent neural network of two-way tree construction, here two-way refers to from top to bottom
Mode from the bottom up, this bi-directional configuration not only transmits the information from leaf node to each node, also to node
Information is transmitted, this is particularly significant for the classification of argument relationship, takes full advantage of the node near tree bottom, top-down knot
Structure sends information near leaf node from top, and can be compatible with the leaf node of different type and quantity, mutually similar
The child node of type shares weight matrix in LSTM unit.The n dimension of t-th of node is calculated in LSTM unit according to following formula
Vector:
ht=ot⊙tanh(ct)
Wherein, m () is mapping function, and C (t) is the child node of t-th of node, and i is shared parameter.
The relationship between two target words pair is indicated using shortest path structure (SPTree), it is for capturing target word
Independent path between.We are stacked on analytic sheaf is relied on sequence layer, by the letter of text sequence and dependence analytic tree
Breath is merged into output, relies on the LSTM input of t-th of word of analytic sheaf are as follows:It is hidden in catenation sequence layer
Hide unit stWith argument relational dependence typeAnd the entity of argument component indicates
All argument physical components identified according to sequence layer, to the last one word of each argument physical components
All situations are arranged out, dependence analytic sheaf is then inputted, export this argument reality finally by two layers of neural net layer
The relationship classification of body component combination.When two argument physical components being drawn into be mistake or they between it is not related,
Relationship between them is regarded as negative relationship, therefore the direction of the type of relationship and relationship is indicated into the relationship between argument.
The dependence of each candidate can be expressed as dp=[↑ hpA:↓hp1:↓hp2], wherein ↓ hpAIndicate two argument entity nodes pair
The hidden layer for the minimum father node answered, ↑ hp1With ↑ hp2It is the hidden state vector of two LSTM units, respectively indicates top-down
First and second On Targets point entity components in LSTM-RNN.
Similar with the Entity recognition of argument type, we realize two layers of neural network, it includes the hidden of a n dimension
Hide layer h(r)With the output layer of a Softmax:
Wherein, W is weight matrix, and b is bias vector.
The LSTM-RNNs of tree construction is superimposed upon on sequence layer, the input d of argument relationship classification is constructedp, at this point,
Sequence layer to rely on analytic sheaf input be do not have it is directive, in order to make full use of argument entity information and solve input dpNothing
To the problem of.The average value of the hidden state vector of each argument physical components is connected to d from sequence layerpTo carry out argument pass
The classification of system obtains following formula:
Wherein, Up1And Up2It is the index of first and second argument entity set of words.
The LSTM-RNNs of tree construction is superimposed upon on sequence layer, the input d of argument relationship classification is constructedp, at this point,
Sequence layer to rely on analytic sheaf input be do not have it is directive, in order to make full use of argument entity information and solve input dpNothing
To the problem of.The average value of the hidden state vector of each argument physical components is connected to d from sequence layerpTo carry out argument pass
The classification of system.
In addition, considering the direction between two argument component entities from left to right and from right to left simultaneously, it is in prediction
Two labels are assigned with to according to direction to each word, when the both direction label of prediction is inconsistent, selection has higher
Final result of the relationship of confidence level as output.Finally training output obtains the label of argument relationship.
The foregoing is merely presently preferred embodiments of the present invention, all equivalent changes done according to scope of the present invention patent with
Modification, is all covered by the present invention.
Claims (8)
1. a kind of based on the joint debate digging system for relying on analytic tree characterized by comprising
One data preprocessing module, for being pre-processed to data;
One text is embedded in module, for from extracted in the text of input word, character, part of speech, between argument dependence and
The vector of argument type indicates;
One sequential coding module, using the contextual information of two-way length Memory Neural Networks learning text in short-term, for completing to discuss
The task of point border detection and argument Relation extraction;
One relies on analytic tree module, relies on analytic tree by building, for finding shortest path warp in two argument component entities;
Label output module is excavated in one debate, and the Tag Estimation work of three tasks, the type of argument are excavated for completing debate
The relational tags of label and argument.
2. according to claim 1 based on the joint debate digging system for relying on analytic tree, it is characterised in that: the data
Preprocessing module carries out pretreatment to data and specifically includes:
(1) web page interlinkage in document, spcial character, punctuation mark are removed;
(2) word segmentation processing is carried out to document;
(3) stem reduction treatment is carried out to English data;
(4) stop words for including in data set is filtered out according to the deactivated vocabulary of Chinese and English respectively.
3. according to claim 1 based on the joint debate digging system for relying on analytic tree, it is characterised in that: the text
It is embedded in module and uses depth convolutional neural networks.
4. -3 any a kind of analytic method based on the joint debate digging system for relying on analytic tree according to claim 1,
Characterized by comprising the following steps:
Step S1: opinion distortion document to be excavated is input to data preprocessing module and is pre-processed, what is obtained is pretreated
Text, and input text insertion module;
Step S2: text be embedded in module using depth convolutional neural networks to pretreated Text Feature Extraction word, character, part of speech,
The vector of dependence and argument type indicates between argument, and list entries coding module;
Step S3: list entries coding module is embedded in the text data that module inputs according to text, uses two-way long short-term memory
The contextual information of neural network learning text data completes argument border detection and argument Relation extraction, obtains argument component
The type label of entity;
Step S4: dependence is closed according to relying between the type label and argument of argument component entity, relies on analytic tree module
Analytic tree is relied on by building, training obtains the label of argument relationship;
Step S5: debate excavates label output module and exports the relational tags of the type label of obtained argument and argument.
5. a kind of analytic method based on the joint debate digging system for relying on analytic tree according to claim 4, special
Sign is: the step S2 specifically:
Step S21: the input of depth convolutional neural networks is pretreated text sequence x=[x1,x2,...,xn], according to text
The sequence of word in sentence, every a line are all the words indicated by d dimensional vector, and CNN output is sequence C=[c1,c2,...,
cn,], C indicates to input the feature of each word, and n indicates the maximum length of list entries;
Step S22: the convolution kernel W ∈ R for the use of one width of narrow convolution sum being k between x(d×k), and willWithThe head and tail portion of sequence are filled into as filling vector;
Step S23: Text Feature Extraction word V is exported respectivelyw, character Vc, part of speech Vp, dependence V between argumentdAnd argument type Ve
Vector indicate, and in list entries coding module.
6. a kind of analytic method based on the joint debate digging system for relying on analytic tree according to claim 4, special
Sign is: the step S3 specifically:
Step S31: the input of sequential coding layer is the sharing feature parameter vector of text embeding layer output, including word Vw, character Vc、
Part of speech Vp, for learning text contextual information and identify argument component entity;
Step S32: one two-way LSTM of building calculates and obtains sentence vector, each LSTM unit is in t-th of word by a n
Dimensional vector composition, comprising: an input gate it, a forgetting door ft, an out gate ot, a memory unit ctAnd one
Hidden unit ht, the vector input of one n dimension of each LSTM unit reception, previous hidden state is ht-1, previous memory list
Member is ct-1;
Undated parameter according to the following formula:
it=σ (W(i)xt+I(i)ht-1+bi)
ft=σ (W(f)xt+I(f)ht-1+bf)
ot=σ (W(o)xt+I(o)ht-1+bo)
ut=tanh (W(u)xt+I(u)ht-1+bu)
ct=it⊙ut+ft⊙ct-1
ht=ot⊙tanh(ct)
Wherein, σ indicates logistic activation primitive, and ⊙ indicates the dot product of vector, and W and I indicate weight matrix, b indicate deviation to
Amount, input of the LSTM unit on t-th of word are the word V of t-th of wordt w, character Vt cWith part of speech Vt pConnection vectorBy the hidden unit of two reversed LSTMWithIt is connected asAs output;
Step S32: upper BIO label is marked to each word of input sentence, its argument type is then marked again, is formed
The form of " BIO- argument type ";
Step S33: one two layers of neural network being made of DenseNet and Softmax of building:
Wherein, W is weight matrix, and b is bias vector;
Step S34: by stWith the vector e of previous wordi-1As input, Softmax is input to by one layer of neural network later
Layer obtains the type label of argument component entity, obtains output and is mapped as vector ei。
7. a kind of analytic method based on the joint debate digging system for relying on analytic tree according to claim 6, special
Sign is: the step S4 specifically:
Step S41: by the type label e of argument component entityiWith the dependence V from text embeding layereInput dependence parsing
Layer;
Step S42: the LSTM combination recurrent neural network of two-way tree construction is constructed, and is counted in LSTM unit by following formula
Calculate the n-dimensional vector of t-th of node:
ht=ot⊙tanh(ct)
Wherein, m () is mapping function, and C (t) is the child node of t-th of node, and i is shared parameter;
The relationship between two target words pair is indicated using shortest path structure, it is used to capture the dependence between target word pair
Path will rely on analytic sheaf and be stacked on sequence layer, text sequence and the information for relying on analytic tree are merged into output, according to
Rely the LSTM input of t-th of word of analytic sheaf are as follows:Hidden unit s in catenation sequence layertIt is closed with argument
System relies on typeAnd the entity of argument component indicates
Step S43: the direction of the type of relationship and relationship is indicated to the relationship between argument, the dependence of each candidate can
To be expressed as dp=[↑ hpA:↓hp1:↓hp2], wherein ↓ hpAIndicate the implicit of the corresponding minimum father node of two argument entity nodes
Layer, ↑ hp1With ↑ hp2It is the hidden state vector of two LSTM units, respectively indicates first in top-down LSTM-RNN and
Two objects argument physical components;
Step S44: one two layers of neural network of setting, it includes the hidden layer h of n dimension(r)With the output of a Softmax
Layer:
Wherein, W is weight matrix, and b is bias vector;
The LSTM-RNNs of tree construction is superimposed upon on sequence layer, the input d of argument relationship classification is constructedp, by each argument
The average value of the hidden state vector of physical components is connected to d from sequence layerpThe classification of argument relationship is carried out, obtains following public affairs
Formula:
Wherein, Up1And Up2It is the index of first and second argument entity set of words;
Step S45: being to be assigned with two labels to according to direction to each word in prediction, when prediction both direction label not
When consistent, select the relationship with high confidence as the final result of output, and training output obtains the mark of argument relationship
Label.
8. a kind of analytic method based on the joint debate digging system for relying on analytic tree according to claim 7, special
Sign is: the two-way mode referred to from top to bottom and from the bottom up of the two-way tree construction, not only comes to the transmitting of each node
From the information of leaf node, information also is transmitted to node.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910034772.XA CN109710946A (en) | 2019-01-15 | 2019-01-15 | A kind of joint debate digging system and method based on dependence analytic tree |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910034772.XA CN109710946A (en) | 2019-01-15 | 2019-01-15 | A kind of joint debate digging system and method based on dependence analytic tree |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109710946A true CN109710946A (en) | 2019-05-03 |
Family
ID=66261394
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910034772.XA Pending CN109710946A (en) | 2019-01-15 | 2019-01-15 | A kind of joint debate digging system and method based on dependence analytic tree |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109710946A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110941700A (en) * | 2019-11-22 | 2020-03-31 | 福州大学 | Multi-task joint learning-based argument mining system and working method thereof |
CN111104797A (en) * | 2019-12-17 | 2020-05-05 | 南开大学 | Paper network representation learning method based on dual sequence-to-sequence generation |
CN112116095A (en) * | 2019-06-19 | 2020-12-22 | 北京搜狗科技发展有限公司 | Method and related device for training multi-task learning model |
CN112487812A (en) * | 2020-10-21 | 2021-03-12 | 上海旻浦科技有限公司 | Nested entity identification method and system based on boundary identification |
CN112732994A (en) * | 2021-01-07 | 2021-04-30 | 上海携宁计算机科技股份有限公司 | Method, device and equipment for extracting webpage information and storage medium |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20100235308A1 (en) * | 2009-03-11 | 2010-09-16 | Shingo Takamatsu | Text analysis device and method and program |
CN107220300A (en) * | 2017-05-05 | 2017-09-29 | 平安科技(深圳)有限公司 | Information mining method, electronic installation and readable storage medium storing program for executing |
CN107944014A (en) * | 2017-12-11 | 2018-04-20 | 河海大学 | A kind of Chinese text sentiment analysis method based on deep learning |
CN108363695A (en) * | 2018-02-23 | 2018-08-03 | 西南交通大学 | A kind of user comment attribute extraction method based on bidirectional dependency syntax tree characterization |
-
2019
- 2019-01-15 CN CN201910034772.XA patent/CN109710946A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20100235308A1 (en) * | 2009-03-11 | 2010-09-16 | Shingo Takamatsu | Text analysis device and method and program |
CN107220300A (en) * | 2017-05-05 | 2017-09-29 | 平安科技(深圳)有限公司 | Information mining method, electronic installation and readable storage medium storing program for executing |
CN107944014A (en) * | 2017-12-11 | 2018-04-20 | 河海大学 | A kind of Chinese text sentiment analysis method based on deep learning |
CN108363695A (en) * | 2018-02-23 | 2018-08-03 | 西南交通大学 | A kind of user comment attribute extraction method based on bidirectional dependency syntax tree characterization |
Non-Patent Citations (3)
Title |
---|
MAKOTO MIWA等: "End-to-End Relation Extraction using LSTMs on Sequences and Tree Structures", 《HTTP://ARXIV.ORG/ABS/1601.00770》 * |
STEFFEN EGER等: "Neural End-to-End Learning for Computational Argumentation Mining", 《HTTP://ARXIV.ORG/ABS/1704.06104》 * |
廖祥文等: "基于多任务迭代学习的论辩挖掘方法", 《计算机学报》 * |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112116095A (en) * | 2019-06-19 | 2020-12-22 | 北京搜狗科技发展有限公司 | Method and related device for training multi-task learning model |
CN110941700A (en) * | 2019-11-22 | 2020-03-31 | 福州大学 | Multi-task joint learning-based argument mining system and working method thereof |
CN110941700B (en) * | 2019-11-22 | 2022-08-09 | 福州大学 | Multi-task joint learning-based argument mining system and working method thereof |
CN111104797A (en) * | 2019-12-17 | 2020-05-05 | 南开大学 | Paper network representation learning method based on dual sequence-to-sequence generation |
CN111104797B (en) * | 2019-12-17 | 2023-05-02 | 南开大学 | Dual-based sequence-to-sequence generation paper network representation learning method |
CN112487812A (en) * | 2020-10-21 | 2021-03-12 | 上海旻浦科技有限公司 | Nested entity identification method and system based on boundary identification |
CN112487812B (en) * | 2020-10-21 | 2021-07-06 | 上海旻浦科技有限公司 | Nested entity identification method and system based on boundary identification |
CN112732994A (en) * | 2021-01-07 | 2021-04-30 | 上海携宁计算机科技股份有限公司 | Method, device and equipment for extracting webpage information and storage medium |
CN112732994B (en) * | 2021-01-07 | 2022-01-28 | 上海携宁计算机科技股份有限公司 | Method, device and equipment for extracting webpage information and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109710946A (en) | A kind of joint debate digging system and method based on dependence analytic tree | |
WO2021147726A1 (en) | Information extraction method and apparatus, electronic device and storage medium | |
CN110334339B (en) | Sequence labeling model and labeling method based on position perception self-attention mechanism | |
CN110196913A (en) | Multiple entity relationship joint abstracting method and device based on text generation formula | |
CN108846017A (en) | The end-to-end classification method of extensive newsletter archive based on Bi-GRU and word vector | |
CN109800437A (en) | A kind of name entity recognition method based on Fusion Features | |
Sun et al. | Deep LSTM networks for online Chinese handwriting recognition | |
CN108363695A (en) | A kind of user comment attribute extraction method based on bidirectional dependency syntax tree characterization | |
CN109960728A (en) | A kind of open field conferencing information name entity recognition method and system | |
CN112966525B (en) | Law field event extraction method based on pre-training model and convolutional neural network algorithm | |
CN114065702A (en) | Event detection method fusing entity relationship and event element | |
CN110134950A (en) | A kind of text auto-collation that words combines | |
CN113836992A (en) | Method for identifying label, method, device and equipment for training label identification model | |
CN111145914B (en) | Method and device for determining text entity of lung cancer clinical disease seed bank | |
CN116245097A (en) | Method for training entity recognition model, entity recognition method and corresponding device | |
Yang et al. | Iterative class prototype calibration for transductive zero-shot learning | |
CN110941700B (en) | Multi-task joint learning-based argument mining system and working method thereof | |
CN113780059A (en) | Continuous sign language identification method based on multiple feature points | |
CN112699685A (en) | Named entity recognition method based on label-guided word fusion | |
Zhao et al. | Domain adaptation with feature and label adversarial networks | |
CN114881038B (en) | Chinese entity and relation extraction method and device based on span and attention mechanism | |
CN116484852A (en) | Chinese patent entity relationship joint extraction method based on relationship diagram attention network | |
Shirghasemi et al. | The impact of active learning algorithm on a cross-lingual model in a Persian sentiment task | |
US20230153533A1 (en) | Pre-training techniques for entity extraction in low resource domains | |
CN115730232A (en) | Topic-correlation-based heterogeneous graph neural network cross-language text classification 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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190503 |
|
RJ01 | Rejection of invention patent application after publication |