CN115577112B - Event extraction method and system based on type perception gated attention mechanism - Google Patents

Event extraction method and system based on type perception gated attention mechanism Download PDF

Info

Publication number
CN115577112B
CN115577112B CN202211576463.3A CN202211576463A CN115577112B CN 115577112 B CN115577112 B CN 115577112B CN 202211576463 A CN202211576463 A CN 202211576463A CN 115577112 B CN115577112 B CN 115577112B
Authority
CN
China
Prior art keywords
event
trigger word
word
text
argument
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.)
Active
Application number
CN202211576463.3A
Other languages
Chinese (zh)
Other versions
CN115577112A (en
Inventor
朱婷婷
杨瀚
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Chengdu Sobey Digital Technology Co Ltd
Original Assignee
Chengdu Sobey Digital Technology Co Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Chengdu Sobey Digital Technology Co Ltd filed Critical Chengdu Sobey Digital Technology Co Ltd
Priority to CN202211576463.3A priority Critical patent/CN115577112B/en
Publication of CN115577112A publication Critical patent/CN115577112A/en
Application granted granted Critical
Publication of CN115577112B publication Critical patent/CN115577112B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06F16/35—Clustering; Classification
    • G06F16/353—Clustering; Classification into predefined classes
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06F16/3331—Query processing
    • G06F16/334—Query execution
    • G06F16/3346—Query execution using probabilistic model
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Databases & Information Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Evolutionary Computation (AREA)
  • Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Machine Translation (AREA)

Abstract

本发明涉及信息抽取技术领域，公开了一种基于类型感知门控注意力机制的事件抽取方法及系统，该事件抽取方法，利用门控信息指导，使不同事件类别下有不同信息流向触发词，过滤与触发词无关的噪声。本发明解决了现有技术存在的事件论元抽取准确度低、角色分类效果差等问题。

Figure 202211576463

The invention relates to the technical field of information extraction, and discloses an event extraction method and system based on a type-aware gating attention mechanism. The event extraction method utilizes gating information guidance to make different information flow to trigger words under different event categories. Filter out noise not related to trigger words. The invention solves the problems of low event argument extraction accuracy, poor role classification effect and the like existing in the prior art.

Figure 202211576463

Description

Event extraction method and system based on type perception gated attention mechanism
Technical Field
The invention relates to the technical field of information extraction, in particular to an event extraction method and system based on a type-aware gated attention mechanism.
Background
Event extraction is a fundamental and extremely challenging task in the field of information extraction. Event extraction generally includes two tasks, namely event detection and event argument extraction. More specifically, the event detection task comprises two subtasks of trigger word detection and event classification, and the event argument extraction comprises two subtasks of argument detection and role classification. In recent years, with the development of deep learning, event extraction methods based on deep learning have been improved to some extent, but the difficulty of event extraction has not yet been completely solved.
At present, most event extraction methods focus on solving the overlapping argument scene, but neglect the overlapping trigger word scene and trigger word ambiguity problem. In other words, rather than arguments that may play different roles in different/the same event, the trigger may have multiple event types.
Furthermore, event argument extraction is more difficult than event detection. Many methods attempt to improve the event argument extraction effect by using role information, such as role occurrence frequency (importance), role correlation such as hierarchical conceptual relationship, role grammatical relationship, and the like. The appearance frequency of the roles ignores the mutual relation among the roles, the correlation among other roles needs to be summarized and concluded based on human experience, and the correlation is not applicable to some data, so the extraction effect of event arguments is not greatly improved.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention provides an event extraction method and system based on a type-aware gated attention mechanism, and solves the problems of low accuracy of event argument extraction, poor role classification effect and the like in the prior art.
The technical scheme adopted by the invention for solving the problems is as follows:
an event extraction method based on a type perception gating attention mechanism utilizes gating information guidance to enable different information under different event types to flow to trigger words and filters noise irrelevant to the trigger words.
As a preferred technical scheme, the method comprises the following steps:
s1, text vectorization: mixing the sample
Figure SMS_1
Inputting the result into a text vectorization layer based on a language model to obtain a text vectorization result ≥>
Figure SMS_2
(ii) a Wherein it is present>
Figure SMS_3
Indicates the th in the text>
Figure SMS_4
Word and/or word>
Figure SMS_5
Represents->
Figure SMS_6
Corresponding vectorized result, based on the result of the evaluation>
Figure SMS_7
Representing the vectorized result of the text X, R representing a real number, d representing the dimensions of a vector, R d Representing a d-dimensional real number vector;
s2, event detection: vectorizing text to a result
Figure SMS_8
Inputting an event detection module of a gate control attention mechanism with fusion type perception to complete two subtasks of trigger word detection and event classification;
s3, argument extraction: and (3) for each trigger word under each event type in the results after the trigger word detection and the event classification are finished in the step (S2), completing two subtasks of argument extraction and argument role classification by using an argument extraction module which integrates learnable role interaction parameters.
As a preferred technical solution, the event detection module of the fusion type perception gated attention mechanism in step S2 includes the following sub-modules connected in series: a trigger word extraction layer and a gated attention event classification layer.
As a preferred technical solution, the construction process of the trigger word extraction layer includes the following steps:
s21, firstly, calculating and obtaining the probability that each character in the input text is a starting/ending character of a trigger word according to the following formula:
Figure SMS_9
wherein,
Figure SMS_10
sigmoid is an activation function for a learnable network parameter, SIGMOID>
Figure SMS_11
For the th in the input text>
Figure SMS_12
The probability that a word is the starting character of a trigger word, <' >>
Figure SMS_13
Is the fifth or fifth in the text>
Figure SMS_14
Probability that a word is the ending character of a trigger word;
s22, according to the preset threshold value
Figure SMS_15
、
Figure SMS_16
Filtering the result in S21 to obtain a location set
Figure SMS_17
、
Figure SMS_18
：
Figure SMS_19
；
Figure SMS_20
；
Wherein,
Figure SMS_21
a set of start character positions representing a trigger word, <' > or>
Figure SMS_22
A set of ending character positions representing trigger words;
s23, combining the result of the step S22, and obtaining a trigger word set by using a nearest matching principle
Figure SMS_23
；
Wherein t is a candidate trigger word, s is the position of the starting character of the candidate trigger word t in the text X,
Figure SMS_24
is a set
Figure SMS_25
Is closest to->
Figure SMS_26
The element (b);
the construction process of the gated attention event classification layer comprises the following steps:
s24, in the gating information filtering layer, for each event category
Figure SMS_27
Defining event class semantic vector>
Figure SMS_28
The corresponding gating vector is calculated as follows:
Figure SMS_29
；
wherein,
Figure SMS_30
is an event category->
Figure SMS_31
Lower gated vector, < >>
Figure SMS_32
Learnable weight parameters for a gating unit>
Figure SMS_33
A learnable bias parameter for the gate control unit;
and S25, combining the result in the S24, and filtering the context information by using an element product function under each event category:
Figure SMS_34
；
wherein,
Figure SMS_35
for the th in the input text>
Figure SMS_36
Word corresponding vector->
Figure SMS_37
For first/second in text input after passing through a gating information filter layer>
Figure SMS_38
Word in event category pick>
Figure SMS_39
The corresponding vector after information filtering;
s26, in the attention information fusion layer, the event category is obtained by using the attention calculation function
Figure SMS_40
The lower input text is ^ h or greater>
Figure SMS_41
Word-to-trigger word->
Figure SMS_42
Is greater than or equal to>
Figure SMS_43
；
And S27, combining the obtained importance scores calculated in S26, obtaining a final information aggregation result related to each trigger word under each event category by using the following formula:
Figure SMS_44
；
wherein,
Figure SMS_45
for post-attention information fusion layer event categories>
Figure SMS_46
The information related to the trigger word t is added to form an aggregation vector;
s28, in the event classification layer, in combination with the information aggregation vector related to the trigger word t obtained in step S27, determining the event type to which the trigger word t belongs by using the following formula:
Figure SMS_47
；
wherein,
Figure SMS_48
learnable weight parameter for an event class decision unit>
Figure SMS_49
Bias parameter for event class determination unit, w T Represents a transposition of w;
according to a preset threshold value
Figure SMS_50
All event classes which satisfy the following condition>
Figure SMS_51
All will be determined as event categories to which the trigger word t belongs:
Figure SMS_52
；
finally, the event category of each trigger word t is collected into
Figure SMS_53
。
As a preferred technical solution, in step S26, the attention calculation function formula is as follows:
Figure SMS_54
；
wherein,
Figure SMS_55
for triggering a word>
Figure SMS_56
Is obtained by the following formula:
Figure SMS_57
；
wherein,
Figure SMS_58
a characterization vector representing the beginning character of the trigger word t, be->
Figure SMS_59
A token vector representing an ending character of the trigger word t;
Figure SMS_60
is defined as follows:
Figure SMS_61
；
wherein,
Figure SMS_62
denotes the tan h activation function, V T Means for transposing a V>
Figure SMS_63
Represents a weight, [; (ii) a]Representing the concatenation of the vectors. />
As a preferred technical solution, the construction process of the argument extraction module fused with learnable role interaction parameters in step S3 includes the following steps:
s31, calculating a characteristic vector of the trigger words blended into the context by using the following formula
Figure SMS_64
：
Figure SMS_65
Wherein,
Figure SMS_79
is based on the input text being in the event category>
Figure SMS_68
Lower information filtered corresponding vector->
Figure SMS_72
The calculated mean value is then taken>
Figure SMS_78
Is based on the input text being in the event category>
Figure SMS_81
The corresponding vector which is filtered by the information is selected>
Figure SMS_80
The standard deviation of the result of the calculation, device for selecting or keeping>
Figure SMS_82
Indicates the ^ th or greater in the input text>
Figure SMS_71
Word in event category pick>
Figure SMS_76
The vector after information filtering and fusion of the trigger word t information is combined>
Figure SMS_66
、
Figure SMS_74
Represents an extension parameter and a translation parameter, respectively>
Figure SMS_69
、
Figure SMS_75
Respectively represent a value for calculating->
Figure SMS_70
Is selected, the weight parameter, the bias parameter, -of the linear layer of (4)>
Figure SMS_77
、
Figure SMS_67
Respectively indicate for calculation &>
Figure SMS_73
The weight parameter and the bias parameter of the linear layer;
s32, respectively calculating each character in the input text as each role event category by using the following formula
Figure SMS_83
Probability size of start character/end character of argument:
Figure SMS_84
wherein,
Figure SMS_87
indicates that the entered text is up to->
Figure SMS_91
Word is in event category->
Figure SMS_94
Is triggered by the trigger word t->
Figure SMS_86
Probability value of the start character of the argument below, <>
Figure SMS_90
Indicates that the entered text is up to->
Figure SMS_93
Word is in event category->
Figure SMS_96
Is triggered by the trigger word t->
Figure SMS_85
Probability value of the end character of the argument below, <>
Figure SMS_89
、
Figure SMS_92
Represents a weight parameter, <' > is selected>
Figure SMS_95
、
Figure SMS_88
Representing a bias parameter;
s33, defining a character interaction matrix capable of learning
Figure SMS_97
And designing the following decision function:
Figure SMS_98
；
wherein,
Figure SMS_99
indicates an event category pick>
Figure SMS_100
An underlying indicator function->
Figure SMS_101
、
Figure SMS_102
A weight parameter and a bias parameter representing a linear layer of the first layer, -a>
Figure SMS_103
Representing a bias parameter of the second layer linear layer;
will be provided with
Figure SMS_104
As the weight, in combination with the calculation result in the weight correction step S32:
Figure SMS_105
；/>
Figure SMS_106
；
after training, the judgment function can learn the mutual relation between the event roles and the mutual relation between the roles at the same time;
wherein,
Figure SMS_108
for entering text ^ h>
Figure SMS_110
Word is event category->
Figure SMS_114
Is triggered by the trigger word t->
Figure SMS_109
The final probability value of the beginning character of the argument below, based on the current location of the argument below>
Figure SMS_111
First ^ for inputting text>
Figure SMS_113
Word is event category->
Figure SMS_116
Is triggered by the trigger word t->
Figure SMS_107
The final probability value of the end character of the argument below, based on the value of the argument below>
Figure SMS_112
Is an event category->
Figure SMS_115
Down character>
Figure SMS_117
The weight of (c);
s34, according to the preset threshold value
Figure SMS_118
、
Figure SMS_119
Filtering the results in S33Thereby obtaining a set of positions
Figure SMS_120
、
Figure SMS_121
：
Figure SMS_122
；
Figure SMS_123
；
Wherein,
Figure SMS_124
indicates that a character is pickand place>
Figure SMS_125
The set of start character positions of the lower argument, <' >>
Figure SMS_126
Indicates that a character is pickand place>
Figure SMS_127
The set of ending character positions of the lower argument;
s35, combining the result of the step S34, and obtaining the role by using the nearest matching principle
Figure SMS_128
Argument set of
Figure SMS_129
(ii) a Wherein it is present>
Figure SMS_130
Is argument>
Figure SMS_131
The position of the start character of (b) in the text X, device for combining or screening>
Figure SMS_132
As a set +>
Figure SMS_133
Is closest to->
Figure SMS_134
Of (2) is used.
As a preferred solution, the loss function of the event extraction method is as follows:
Figure SMS_135
；
wherein,
Figure SMS_139
representing input text>
Figure SMS_143
Event type of the middle trigger word t, < >>
Figure SMS_147
Representing input text
Figure SMS_138
The trigger word t with medium event type c has a role of ≥>
Figure SMS_142
Argument of (D), D denotes all input samples, based on the value of the argument, and->
Figure SMS_144
Indicates that in sample x the trigger word is t and the event type is c and argument of character r is +>
Figure SMS_149
In the case of>
Figure SMS_136
Represents the probability that the trigger word in sample x is t, <' >>
Figure SMS_141
Represents a trigger word in a sample x->
Figure SMS_145
Has a probability of c, is greater than>
Figure SMS_148
Indicating in sample x that the event category is c and the trigger word is &>
Figure SMS_137
Argument of time character r is ^ 5>
Figure SMS_140
Is greater than or equal to>
Figure SMS_146
Representing all events in sample x.
As a preferred technical solution, in step S1, the language model is a BERT model.
A text vectorization module: for mixing the sample
Figure SMS_150
Inputting the result into a text vectorization layer based on a language model to obtain a text vectorization result ≥>
Figure SMS_151
(ii) a Wherein,
Figure SMS_152
indicates the th in the text>
Figure SMS_153
Word and/or word>
Figure SMS_154
Represents->
Figure SMS_155
Corresponding vectorized result, is>
Figure SMS_156
Representing vectorization results of a text X, R representing a real number, d representing a dimension of a vector, R d Representing a d-dimensional real number vector; />
An event detection module: to vectorize the text into a result
Figure SMS_157
Inputting an event detection module of a gate control attention mechanism with fusion type perception to complete two subtasks of trigger word detection and event classification;
argument extraction module: and the argument extraction module which integrates learnable role interaction parameters is used for completing two subtasks of argument extraction and argument role classification for each trigger word under each event type in the results after the event detection module completes trigger word detection and event classification.
Compared with the prior art, the invention has the following beneficial effects:
according to the invention, different information flows to the trigger words under different event types are guided by using the gating information, noise irrelevant to the trigger words is effectively filtered, and other relevant information is effectively integrated, so that the ambiguity problem of the trigger words under the event extraction scene containing the overlapped trigger words is better eliminated, the overlapped trigger word scene is dealt with, and the event classification effect is improved; meanwhile, the method considers the role co-occurrence relation ignored by people, and further improves the effect of argument extraction and role classification tasks by introducing learnable role interaction parameters to model the role co-occurrence relation.
Drawings
FIG. 1 is a schematic diagram illustrating steps of an event extraction method based on a type-aware gated attention mechanism according to the present invention;
fig. 2 is a schematic diagram of a model structure of an event detection module of a fusion type perception gated attention mechanism according to an embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to examples and drawings, but the embodiments of the present invention are not limited thereto.
The invention provides an event extraction method based on a type perception gated attention mechanism, which designs the type perception gated attention mechanism, and is different from a pure attention mechanism, the method utilizes gated information guidance to enable different information to flow to trigger words under different event types, effectively filters noise irrelevant to the trigger words and effectively integrates other related information, thereby better eliminating the ambiguity problem of the trigger words under an event extraction scene containing overlapped trigger words, coping with the overlapped trigger word scene and improving the event classification effect; meanwhile, the method considers the role co-occurrence relation ignored by everybody, and further improves the effect of argument extraction and role classification tasks by introducing learnable role interaction parameters to model the role co-occurrence relation.
Example 1
As shown in fig. 1, an event extraction method based on a type-aware gated attention mechanism includes the steps of:
s1, mixing the sample
Figure SMS_158
Inputting the text vector into a text vectorization layer based on a language model such as BERT to obtain a text vectorization result->
Figure SMS_159
. Wherein
Figure SMS_160
Indicates the th in the text>
Figure SMS_161
Word and/or word>
Figure SMS_162
The corresponding vectorization result is obtained;
s2, vectorizing the text into a result
Figure SMS_163
Inputting an event detection module of a gate control attention mechanism with fusion type perception to complete two subtasks of trigger word detection and event classification;
s3, for each trigger word belonging to each event type in the result of S2, completing two subtasks of argument extraction and argument role classification by using an argument extraction module which integrates learnable role interaction parameters;
example 2
On the basis of embodiment 1, as shown in fig. 2, the event detection module of the fusion type perception gated attention mechanism in step S2 includes the following sub-modules in a serial order: a trigger word extraction layer and a gated attention event classification layer.
The construction process of the trigger word extraction layer comprises the following steps:
s21, firstly, calculating and obtaining the probability that each character in the input text is a starting/ending character of the trigger word according to the following formula:
Figure SMS_164
wherein,
Figure SMS_165
sigmoid is an activation function for a learnable network parameter.
Figure SMS_166
Corresponding to the ^ th or greater part of the input text>
Figure SMS_167
Probability that a word is the starting character of a trigger word, based on the word's word score>
Figure SMS_168
Then corresponds to the ^ th in the text>
Figure SMS_169
Probability that a word is the ending character of a trigger word;
s22, according to the preset threshold value
Figure SMS_170
、
Figure SMS_171
Filtering the result in S21 to obtain a location set
Figure SMS_172
、
Figure SMS_173
：
Figure SMS_174
Figure SMS_175
Here, the location sets
Figure SMS_176
、
Figure SMS_177
Respectively representing a starting character position set and an ending character position set of the trigger word;
s23, combining the result of the step S22, and obtaining a trigger word set by using a nearest matching principle
Figure SMS_178
Here +>
Figure SMS_179
For the position of the start character of the candidate trigger word t in the text X, ->
Figure SMS_180
As a set +>
Figure SMS_181
Is closest to->
Figure SMS_182
Of (2) is used.
The construction process of the gated attention event classification layer comprises the following steps:
s24, in the gating information filtering layer, for each event category
Figure SMS_183
Defining event class semantic vector>
Figure SMS_184
The corresponding gating vector is calculated as follows:
Figure SMS_185
here, the
Figure SMS_186
For an event category>
Figure SMS_187
Lower gated vector, < >>
Figure SMS_188
Is a learnable weight parameter of the gate control unit>
Figure SMS_189
A learnable bias parameter for the gate control unit;
and S25, combining the result in the S24, and filtering the context information by using an element-wise product (element-wise product) under each event category:
Figure SMS_190
here, ,
Figure SMS_191
for the first ^ in the input text>
Figure SMS_192
Word corresponding vector->
Figure SMS_193
Is the first or second based on the input text after passing through the gating information filtering layer>
Figure SMS_194
Word in event category>
Figure SMS_195
The corresponding vector after information filtering;
s26, in the notesIn the intention information fusion layer, after the calculation described in step S25, the event type is obtained by using the attention calculation function designed as follows
Figure SMS_196
On/in text>
Figure SMS_197
Word-to-trigger word->
Figure SMS_198
Is greater than or equal to>
Figure SMS_199
：
Figure SMS_200
Here, ,
Figure SMS_201
for triggering a word>
Figure SMS_202
Is obtained by the following formula:
Figure SMS_203
in addition to this, the present invention is,
Figure SMS_204
is defined as follows:
Figure SMS_205
；
and S27, combining the obtained importance scores calculated in S26, obtaining a final information aggregation result related to each trigger word under each event category by using the following formula:
Figure SMS_206
here, ,
Figure SMS_207
for an event category after the attention information fusion layer>
Figure SMS_208
Down and trigger word->
Figure SMS_209
A related information aggregation vector;
and S28, in the event classification layer, combining the information aggregation vector related to the trigger word t obtained in the step S27, and determining which event types the trigger word t belongs to by using the following formula:
Figure SMS_210
here, ,
Figure SMS_211
learnable weight parameter for an event class decision unit>
Figure SMS_212
Bias parameter for event class determination unit, w T Represents a transposition of w; in particular according to a predetermined threshold value>
Figure SMS_213
All event classes that satisfy the following condition>
Figure SMS_214
All will be determined as event categories to which the trigger word t belongs:
Figure SMS_215
i.e. finally the set of event classes per trigger word t is
Figure SMS_216
。
Example 3
On the basis of the embodiments 1 and 2, the construction process of the argument extraction module fusing the learnable role interaction parameters in the step S3 includes the following steps:
s31, in order to better extract the event type as
Figure SMS_217
All arguments associated with a given trigger word t are first blended into the trigger word characterization vector @incontext using the following formula>
Figure SMS_218
：
Figure SMS_219
Here, ,
Figure SMS_220
、
Figure SMS_221
is based on the input text being in the event category>
Figure SMS_222
The corresponding vector which is filtered by the information is selected>
Figure SMS_223
Calculating the obtained mean value and standard deviation;
s32, respectively calculating each character in the input text as each role event category by using the following formula
Figure SMS_224
Probability size of start character/end character of argument:
Figure SMS_225
here, the number of the first and second electrodes,
Figure SMS_226
、
Figure SMS_227
respectively represent the input text ^ h>
Figure SMS_228
Word is event category->
Figure SMS_229
Is triggered by the trigger word t->
Figure SMS_230
Probability value of starting character/ending character of argument under;
s33, defining a character interaction matrix capable of learning
Figure SMS_231
And designing the following decision function:
Figure SMS_232
using the calculation result as a weight, and combining the calculation result in the weight correction step S32:
Figure SMS_233
；
Figure SMS_234
；
after training, the judgment function can learn the mutual relation between the event roles and the mutual relation between the roles at the same time.
Figure SMS_236
、
Figure SMS_239
For entering text ^ h>
Figure SMS_241
Word is event category->
Figure SMS_237
Is triggered by the trigger word t->
Figure SMS_238
Final probability values for the beginning/ending characters of the argument below. Here->
Figure SMS_240
Is an event category->
Figure SMS_242
Lower role->
Figure SMS_235
The weight of (c).
S34, according to the preset threshold value
Figure SMS_243
、
Figure SMS_244
Filtering the result in S33 to obtain a set of locations
Figure SMS_245
、
Figure SMS_246
：
Figure SMS_247
；
Figure SMS_248
；/>
Here, the position sets
Figure SMS_249
、
Figure SMS_250
Respectively represent a role->
Figure SMS_251
A starting character position set and an ending character position set of a lower argument;
s35, combining the result of the step S34, and obtaining the role by using the nearest matching principle
Figure SMS_252
Argument set of
Figure SMS_253
Here +>
Figure SMS_254
Is argument>
Figure SMS_255
The position of the start character of (b) in the text X, device for selecting or keeping>
Figure SMS_256
Is set->
Figure SMS_257
Is closest to->
Figure SMS_258
Of (2) is used.
Example 4
On the basis of examples 1, 2 and 3, the loss function of the event extraction method is as follows:
Figure SMS_259
here, ,
Figure SMS_260
indicates that the input text pick>
Figure SMS_261
Event type in which the word t is triggered>
Figure SMS_262
Indicates that the input text pick>
Figure SMS_263
Medium event type is->
Figure SMS_264
Has the role of ≥>
Figure SMS_265
The argument of (1).
As described above, the present invention can be preferably implemented.
All features disclosed in all embodiments in this specification, or all methods or process steps implicitly disclosed, may be combined and/or expanded, or substituted, in any way, except for mutually exclusive features and/or steps.
While the invention has been described in connection with what is presently considered to be the most practical and preferred embodiment, it is to be understood that the invention is not to be limited to the disclosed embodiment, but on the contrary, is intended to cover various modifications, equivalent arrangements, and alternatives falling within the spirit and scope of the invention as defined by the appended claims.

Claims (6)

1. An event extraction method based on a type perception gating attention mechanism is characterized in that different information flows to a trigger word under different event types by using gating information guidance, and noise irrelevant to the trigger word is filtered;
the method comprises the following steps:
s1, text vectorization: mixing the sample
Figure QLYQS_1
Inputting the result into a text vectorization layer based on a language model to obtain a text vectorization result ≥>
Figure QLYQS_2
(ii) a Wherein +>
Figure QLYQS_3
Indicates the th in the text>
Figure QLYQS_4
Word in and out apparatus>
Figure QLYQS_5
Represents->
Figure QLYQS_6
Corresponding vectorized result, is>
Figure QLYQS_7
Representing the vectorized result of the text X, R representing a real number, d representing the dimensions of a vector, R d Representing a d-dimensional real number vector;
s2, event detection: vectorizing text into results
Figure QLYQS_8
Inputting an event detection module of a gate control attention mechanism with fusion type perception to complete two subtasks of trigger word detection and event classification;
s3, argument extraction: for each trigger word under each event type in the results after the trigger word detection and the event classification are completed in the step S2, an argument extraction module which integrates learnable role interaction parameters is utilized to complete two subtasks of argument extraction and argument role classification;
the event detection module of the fusion type perception gated attention mechanism in the step S2 comprises the following serially connected sub-modules: a trigger word extraction layer and a gate control attention event classification layer;
the construction process of the trigger word extraction layer comprises the following steps:
s21, firstly, calculating and obtaining the probability that each character in the input text is a starting/ending character of the trigger word according to the following formula:
Figure QLYQS_9
wherein,
Figure QLYQS_10
sigmoid is an activation function for learnable network parameters>
Figure QLYQS_11
For the first ^ in the input text>
Figure QLYQS_12
The probability that a word is the starting character of a trigger word, <' >>
Figure QLYQS_13
Is the fifth or fifth in the text>
Figure QLYQS_14
The probability that a word is the ending character of a trigger word;
s22, according to the preset threshold value
Figure QLYQS_15
、
Figure QLYQS_16
The result in S21 is filtered, so that a position set @isobtained>
Figure QLYQS_17
、
Figure QLYQS_18
：
Figure QLYQS_19
；
Figure QLYQS_20
；
Wherein,
Figure QLYQS_21
set of start character positions, representing a trigger word>
Figure QLYQS_22
A set of ending character positions representing trigger words;
s23, combining the result of the step S22, and obtaining a trigger word set by using a nearest matching principle
Figure QLYQS_23
；
Wherein t is a candidate trigger word, s is the position of the starting character of the candidate trigger word t in the text X,
Figure QLYQS_24
is set->
Figure QLYQS_25
Is closest to->
Figure QLYQS_26
An element of (1); />
The construction process of the gated attention event classification layer comprises the following steps:
s24, in the gating information filtering layer, for each event category
Figure QLYQS_27
Defining event class semantic vector>
Figure QLYQS_28
The corresponding gating vector is calculated as follows:
Figure QLYQS_29
；
wherein,
Figure QLYQS_30
for an event category>
Figure QLYQS_31
Lower gated vector, < >>
Figure QLYQS_32
Is a learnable weight parameter of the gate control unit>
Figure QLYQS_33
A learnable bias parameter for the gate control unit;
and S25, combining the result in the S24, and filtering the context information by using an element product function under each event category:
Figure QLYQS_34
；
wherein,
Figure QLYQS_35
for the th in the input text>
Figure QLYQS_36
Word corresponding vector->
Figure QLYQS_37
Is the first or second based on the input text after passing through the gating information filtering layer>
Figure QLYQS_38
Word in event category pick>
Figure QLYQS_39
The corresponding vector after information filtering;
s26, in the attention information fusion layer, the event type is obtained by using the attention calculation function
Figure QLYQS_40
Enter text into
Figure QLYQS_41
Word-to-trigger word->
Figure QLYQS_42
Importance score of
Figure QLYQS_43
；
And S27, combining the obtained importance scores calculated in S26, obtaining a final information aggregation result related to each trigger word under each event category by using the following formula:
Figure QLYQS_44
；
wherein,
Figure QLYQS_45
for post-attention information fusion layer event categories>
Figure QLYQS_46
Down and trigger word->
Figure QLYQS_47
A related information aggregation vector;
s28, in the event classification layer, in combination with the information aggregation vector related to the trigger word t obtained in step S27, determining the event type to which the trigger word t belongs by using the following formula:
Figure QLYQS_48
；
wherein,
Figure QLYQS_49
learnable weight parameter for an event class decision unit>
Figure QLYQS_50
Bias parameter for event class determination unit, w T Represents a transposition of w;
according to a preset threshold value
Figure QLYQS_51
All ofEvent category £ that satisfies the following condition>
Figure QLYQS_52
All will be determined as event categories to which the trigger word t belongs:
Figure QLYQS_53
；
finally, each trigger word
Figure QLYQS_54
Is/is set as->
Figure QLYQS_55
。
2. The method for extracting events based on the type-aware gated attention mechanism as claimed in claim 1, wherein in step S26, the attention calculation function formula is as follows:
Figure QLYQS_56
；
wherein,
Figure QLYQS_57
is a trigger word>
Figure QLYQS_58
The characterization vector of (2) is obtained by calculating the following formula:
Figure QLYQS_59
；
wherein,
Figure QLYQS_60
a characterization vector representing the beginning character of the trigger word t, be->
Figure QLYQS_61
A token vector representing an ending character of the trigger word t;
Figure QLYQS_62
is defined as follows:
Figure QLYQS_63
；
wherein,
Figure QLYQS_64
denotes the tan h activation function, V T Represents the transposition of a V>
Figure QLYQS_65
Represents a weight, [; (ii) a]Representing the concatenation of the vectors.
3. The method for extracting events based on the type-aware gated attention mechanism according to claim 2, wherein the construction process of the argument extraction module fused with learnable role interaction parameters in step S3 includes the following steps:
s31, calculating a characteristic vector of the trigger words blended into the context by using the following formula
Figure QLYQS_66
：
Figure QLYQS_67
Wherein,
Figure QLYQS_71
is based on the input text in the event category>
Figure QLYQS_70
The corresponding vector which is filtered by the information is selected>
Figure QLYQS_78
The calculated mean value is then taken>
Figure QLYQS_72
Is based on the input text being in the event category>
Figure QLYQS_76
Lower information filtered corresponding vector->
Figure QLYQS_73
The calculated standard deviation is->
Figure QLYQS_75
Indicates the ^ th or greater in the input text>
Figure QLYQS_80
Word in event category pick>
Figure QLYQS_84
The vector after information filtering and fusion of the trigger word t information is selected and matched>
Figure QLYQS_68
、
Figure QLYQS_74
Represents an extension parameter and a translation parameter, respectively>
Figure QLYQS_77
、
Figure QLYQS_82
Respectively represent a value for calculating->
Figure QLYQS_81
Weight parameter, bias parameter, based on the linear layer of (4), based on the number of linear layers in the device>
Figure QLYQS_83
、
Figure QLYQS_69
Respectively indicate that it is used for counting->
Figure QLYQS_79
The weight parameter and the bias parameter of the linear layer;
s32, respectively calculating each character in the input text as each role event category by using the following formula
Figure QLYQS_85
Probability size of start/end character for argument: />
Figure QLYQS_86
Wherein,
Figure QLYQS_89
indicates that the entered text is up to->
Figure QLYQS_92
Word is event category->
Figure QLYQS_94
Is triggered by the trigger word t->
Figure QLYQS_90
Probability value of the start character of the argument below, <>
Figure QLYQS_91
Indicating that input text is on &>
Figure QLYQS_96
Word is in event category->
Figure QLYQS_98
Is triggered by the trigger word t->
Figure QLYQS_87
Probability value of the end character of the argument below, <>
Figure QLYQS_93
、
Figure QLYQS_95
Represents a weight parameter, <' > is selected>
Figure QLYQS_97
、
Figure QLYQS_88
Representing a bias parameter;
s33, defining a character interaction matrix capable of learning
Figure QLYQS_99
And designing the following decision function:
Figure QLYQS_100
；
wherein,
Figure QLYQS_101
indicates an event category pick>
Figure QLYQS_102
An underlying indicator function->
Figure QLYQS_103
、
Figure QLYQS_104
A weight parameter and a bias parameter representing a linear layer of the first layer, -a>
Figure QLYQS_105
Representing a bias parameter of the second layer linear layer;
will be provided with
Figure QLYQS_106
As the weight, in combination with the calculation result in the weight correction step S32:
Figure QLYQS_107
；
Figure QLYQS_108
；
after training, the judgment function can learn the mutual relation between the event roles and the mutual relation between the roles at the same time;
wherein,
Figure QLYQS_110
for entering text ^ h>
Figure QLYQS_113
Word is event category->
Figure QLYQS_116
Trigger word of t role->
Figure QLYQS_111
The final probability value of the beginning character of the argument below, based on the current location of the argument below>
Figure QLYQS_112
For entering text ^ h>
Figure QLYQS_115
Word is event category->
Figure QLYQS_118
Is triggered by the trigger word t->
Figure QLYQS_109
Of the ending character of the argument belowFinal probability value->
Figure QLYQS_114
Is an event category->
Figure QLYQS_117
Lower role->
Figure QLYQS_119
The weight of (c);
s34, according to the preset threshold value
Figure QLYQS_120
、
Figure QLYQS_121
The result in S33 is filtered, so that a position set @isobtained>
Figure QLYQS_122
、
Figure QLYQS_123
：
Figure QLYQS_124
；
Figure QLYQS_125
；
Wherein,
Figure QLYQS_126
indicates that a character is pickand place>
Figure QLYQS_127
The set of start character positions of the lower argument, <' >>
Figure QLYQS_128
Indicates that a character is pickand place>
Figure QLYQS_129
The ending character position set of the lower argument;
s35, combining the result of the step S34, and obtaining the role by using the nearest matching principle
Figure QLYQS_130
Argument set of
Figure QLYQS_131
(ii) a Wherein it is present>
Figure QLYQS_132
Is argument>
Figure QLYQS_133
The position of the starting character in the text X, device for combining or screening>
Figure QLYQS_134
As a set +>
Figure QLYQS_135
Is closest to->
Figure QLYQS_136
The element (c) of (a). />
4. The event extraction method based on the type-aware gated attention mechanism as claimed in claim 3, wherein the loss function of the event extraction method is as follows:
Figure QLYQS_137
；
wherein,
Figure QLYQS_140
indicates that the input text pick>
Figure QLYQS_145
Middle touchEvent type for the asserted t->
Figure QLYQS_149
Indicates that the input text pick>
Figure QLYQS_139
The trigger word t with medium event type c has a role of ≥>
Figure QLYQS_143
Argument of (D), D denotes all input samples, based on the value of the argument, and->
Figure QLYQS_147
Indicates that the trigger is t and the event type is c in sample x and the argument of the role r is ^ H>
Figure QLYQS_151
Is greater than or equal to>
Figure QLYQS_138
Represents the probability that the trigger word in sample x is t, <' >>
Figure QLYQS_142
Represents a trigger word in a sample x->
Figure QLYQS_146
Has a probability of c, is greater than>
Figure QLYQS_150
Indicates that in sample x the event class is c and the trigger word is ≥>
Figure QLYQS_141
Argument of time character r is ^ 5>
Figure QLYQS_144
Is greater than or equal to>
Figure QLYQS_148
Representing what is in sample xThere is an event.
5. The method for extracting events based on the type-aware gated attention mechanism as claimed in any one of claims 1 to 4, wherein in step S1, the language model is a BERT model.
6. An event extraction system based on a type-aware gated attention mechanism, for implementing the event extraction method based on the type-aware gated attention mechanism as claimed in any one of claims 1 to 5, comprising the following modules connected in sequence:
a text vectorization module: for mixing the sample
Figure QLYQS_152
Inputting into a text vectorization layer based on a language model, obtaining a text vectorization result>
Figure QLYQS_153
(ii) a Wherein +>
Figure QLYQS_154
Indicates the th in the text>
Figure QLYQS_155
Word in and out apparatus>
Figure QLYQS_156
Represents->
Figure QLYQS_157
Corresponding vectorized result, is>
Figure QLYQS_158
Representing the vectorized result of the text X, R representing a real number, d representing the dimensions of a vector, R d Representing a d-dimensional real number vector;
an event detection module: to vectorize the text into a result
Figure QLYQS_159
Inputting an event detection module of a gate control attention mechanism with fusion type perception to complete two subtasks of trigger word detection and event classification;
argument extraction module: the argument extraction module which is integrated with learnable role interaction parameters is used for completing two subtasks of argument extraction and argument role classification for each trigger word under each event type in the results of the event detection module after the trigger word detection and the event classification are completed;
the event detection module of the fusion type perception gating attention mechanism comprises the following sub-modules which are connected in series: a trigger word extraction layer and a gated attention event classification layer;
the construction process of the trigger word extraction layer comprises the following steps:
s21, firstly, calculating and obtaining the probability that each character in the input text is a starting/ending character of a trigger word according to the following formula:
Figure QLYQS_160
wherein,
Figure QLYQS_161
sigmoid is an activation function for a learnable network parameter, SIGMOID>
Figure QLYQS_162
For the th in the input text>
Figure QLYQS_163
Probability that a word is the starting character of a trigger word, based on the word's word score>
Figure QLYQS_164
For the fifth in the text>
Figure QLYQS_165
The probability that a word is the ending character of a trigger word;
s22, according to the preset threshold value
Figure QLYQS_166
、
Figure QLYQS_167
The result in S21 is filtered, so that a position set @isobtained>
Figure QLYQS_168
、
Figure QLYQS_169
：
Figure QLYQS_170
；
Figure QLYQS_171
；
Wherein,
Figure QLYQS_172
a set of start character positions representing a trigger word, <' > or>
Figure QLYQS_173
A set of ending character positions representing trigger words;
s23, combining the result of the step S22, and obtaining a trigger word set by using a nearest matching principle
Figure QLYQS_174
；
Wherein t is a candidate trigger word, s is the position of the starting character of the candidate trigger word t in the text X,
Figure QLYQS_175
is set->
Figure QLYQS_176
Is closest to->
Figure QLYQS_177
The element (b);
the construction process of the gated attention event classification layer comprises the following steps:
s24, in the gating information filtering layer, for each event category
Figure QLYQS_178
Defining an event class semantic vector ≥>
Figure QLYQS_179
The corresponding gating vector is calculated as follows:
Figure QLYQS_180
；
wherein,
Figure QLYQS_181
is an event category->
Figure QLYQS_182
A lower gated vector +>
Figure QLYQS_183
Is a learnable weight parameter of the gate control unit>
Figure QLYQS_184
A learnable bias parameter for the gate control unit;
and S25, combining the result in the S24, and filtering the context information by using an element product function under each event category:
Figure QLYQS_185
；
wherein,
Figure QLYQS_186
for the first ^ in the input text>
Figure QLYQS_187
A vector corresponding to a word, <' >>
Figure QLYQS_188
Is the first or second based on the input text after passing through the gating information filtering layer>
Figure QLYQS_189
Word in event category pick>
Figure QLYQS_190
The corresponding vector after information filtering;
s26, in the attention information fusion layer, the event category is obtained by using the attention calculation function
Figure QLYQS_191
Enter text into
Figure QLYQS_192
Word-to-trigger word->
Figure QLYQS_193
Is greater than or equal to>
Figure QLYQS_194
；
And S27, combining the obtained importance scores calculated in S26, obtaining a final information aggregation result related to each trigger word under each event category by using the following formula:
Figure QLYQS_195
；
wherein,
Figure QLYQS_196
for an event category after the attention information fusion layer>
Figure QLYQS_197
The information related to the trigger word t is aggregated to form a vector;
s28, in the event classification layer, in combination with the information aggregation vector related to the trigger word t obtained in step S27, determining the event type to which the trigger word t belongs by using the following formula:
Figure QLYQS_198
；
wherein,
Figure QLYQS_199
learnable weight parameter for an event class decision unit>
Figure QLYQS_200
Bias parameter for event class determination unit, w T Represents a transposition of w;
according to a preset threshold value
Figure QLYQS_201
All event classes that satisfy the following condition>
Figure QLYQS_202
All will be determined as event categories to which the trigger word t belongs:
Figure QLYQS_203
；
finally, the event category of each trigger word t is collected into
Figure QLYQS_204
。/>
CN202211576463.3A 2022-12-09 2022-12-09 Event extraction method and system based on type perception gated attention mechanism Active CN115577112B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211576463.3A CN115577112B (en) 2022-12-09 2022-12-09 Event extraction method and system based on type perception gated attention mechanism

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211576463.3A CN115577112B (en) 2022-12-09 2022-12-09 Event extraction method and system based on type perception gated attention mechanism

Publications (2)

Publication Number Publication Date
CN115577112A CN115577112A (en) 2023-01-06
CN115577112B true CN115577112B (en) 2023-04-18

Family

ID=84589980

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211576463.3A Active CN115577112B (en) 2022-12-09 2022-12-09 Event extraction method and system based on type perception gated attention mechanism

Country Status (1)

Country Link
CN (1) CN115577112B (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112307761A (en) * 2020-11-19 2021-02-02 新华智云科技有限公司 Event extraction method and system based on attention mechanism
CN113705218A (en) * 2021-09-03 2021-11-26 四川大学 Event element gridding extraction method based on character embedding, storage medium and electronic device
CN115238685A (en) * 2022-09-23 2022-10-25 华南理工大学 Combined extraction method for building engineering change events based on position perception

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10565229B2 (en) * 2018-05-24 2020-02-18 People.ai, Inc. Systems and methods for matching electronic activities directly to record objects of systems of record
US10186275B2 (en) * 2017-03-31 2019-01-22 Hong Fu Jin Precision Industry (Shenzhen) Co., Ltd. Sharing method and device for video and audio data presented in interacting fashion
CN109299470B (en) * 2018-11-01 2024-02-09 成都数联铭品科技有限公司 Method and system for extracting trigger words in text bulletin
CN113239142B (en) * 2021-04-26 2022-09-23 昆明理工大学 Trigger-word-free event detection method fused with syntactic information
CN115292568B (en) * 2022-03-02 2023-11-17 内蒙古工业大学 Civil news event extraction method based on joint model
CN114298053B (en) * 2022-03-10 2022-05-24 中国科学院自动化研究所 A joint event extraction system based on the fusion of feature and attention mechanism
CN114648016B (en) * 2022-03-29 2025-02-11 河海大学 An event argument extraction method based on event element interaction and label semantic enhancement
CN115392248A (en) * 2022-06-22 2022-11-25 北京航空航天大学 A Method for Event Extraction Based on Context and Graph Attention
CN115168541B (en) * 2022-07-13 2025-08-26 山西大学 Text event extraction method and system based on frame semantic mapping and type awareness

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112307761A (en) * 2020-11-19 2021-02-02 新华智云科技有限公司 Event extraction method and system based on attention mechanism
CN113705218A (en) * 2021-09-03 2021-11-26 四川大学 Event element gridding extraction method based on character embedding, storage medium and electronic device
CN115238685A (en) * 2022-09-23 2022-10-25 华南理工大学 Combined extraction method for building engineering change events based on position perception

Also Published As

Publication number Publication date
CN115577112A (en) 2023-01-06

Similar Documents

Publication Publication Date Title
CN110598005B (en) Public safety event-oriented multi-source heterogeneous data knowledge graph construction method
Li et al. A model of coherence based on distributed sentence representation
CN108170816A (en) A kind of intelligent vision Question-Answering Model based on deep neural network
CN108052547A (en) Natural language question-answering method and system based on question sentence and knowledge graph structural analysis
CN114265937A (en) Intelligent classification analysis method and system of scientific and technological information, storage medium and server
US20210312133A1 (en) Word vector-based event-driven service matching method
CN103207860A (en) Method and device for extracting entity relationships of public sentiment events
CN109710744A (en) A kind of data matching method, device, equipment and storage medium
Guo et al. Divert more attention to vision-language object tracking
CN115858723B (en) A method and system for query graph generation in complex knowledge base question answering
CN113158673A (en) Single document analysis method and device
CN108846000A (en) A kind of common sense semanteme map construction method and device based on supernode and the common sense complementing method based on connection prediction
CN118821779B (en) A method and device for extracting relations based on entity pairs and shortest dependency paths
CN109117891A (en) It merges social networks and names across the social media account matching process of feature
Fan et al. Adaptive domain-aware representation learning for speech emotion recognition
CN120707944A (en) Clothing compatibility analysis method, equipment and medium based on multimodal feature fusion
Wang et al. A survey of vision and language related multi-modal task
CN121841709A (en) CVE-ATT &amp; CK association fusion RAG construction method for network security scene
CN106570167A (en) Knowledge-integrated subject model-based microblog topic detection method
CN118350379B (en) Method, device and medium for improving accuracy of natural language processing in knowledge system
CN110377753B (en) Relation extraction method and device based on relation trigger word and GRU model
CN119513317A (en) Zero-shot stance detection method and system based on multi-task fine-tuning and knowledge enhancement of debate data
CN115577112A (en) An event extraction method and system based on type-aware gating attention mechanism
CN116522346A (en) Intelligent contract vulnerability detection method based on key data flow graph and abstract semantic graph
CN107886942A (en) A Speech Signal Emotion Recognition Method Based on Locally Penalized Random Spectral Regression

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant