CN114742016A - Chapter-level event extraction method and device based on multi-granularity entity differential composition - Google Patents

Chapter-level event extraction method and device based on multi-granularity entity differential composition Download PDF

Info

Publication number
CN114742016A
CN114742016A CN202210348614.3A CN202210348614A CN114742016A CN 114742016 A CN114742016 A CN 114742016A CN 202210348614 A CN202210348614 A CN 202210348614A CN 114742016 A CN114742016 A CN 114742016A
Authority
CN
China
Prior art keywords
entity
sentence
event
entities
sentences
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202210348614.3A
Other languages
Chinese (zh)
Other versions
CN114742016B (en
Inventor
张虎
张广军
雷登斌
李茹
梁吉业
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.)
Shanxi University
Original Assignee
Shanxi University
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 Shanxi University filed Critical Shanxi University
Priority to CN202210348614.3A priority Critical patent/CN114742016B/en
Publication of CN114742016A publication Critical patent/CN114742016A/en
Application granted granted Critical
Publication of CN114742016B publication Critical patent/CN114742016B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/10—Text processing
    • G06F40/12—Use of codes for handling textual entities
    • G06F40/126—Character encoding
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/2431—Multiple classes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/279—Recognition of textual entities
    • G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
    • G06F40/295—Named entity recognition
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • 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)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Machine Translation (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention relates to the fields of deep learning, natural language processing and the like, and discloses a chapter-level event extraction method and device based on a multi-granularity entity heterogeneous composition, aiming at the problems existing in the existing entity extraction. The method of the invention respectively extracts entities by using context information based on sentences and paragraphs, and fuses entity sets of two granularities based on a multi-granularity entity selection strategy, thereby improving the accuracy of entity extraction. And then, combining the sentences and the screened candidate entities to construct a heterogeneous graph fused with the multi-granularity entities, and obtaining the entities with chapter-level context perception and vectorized representation of the sentences by using a graph volume network, so that the perception capability of the sentences and the entities to the events is improved. And finally, performing multi-label classification on the event types and the event arguments to realize event detection and argument identification.

Description

Chapter-level event extraction method and device based on multi-granularity entity differential composition
Technical Field
The invention relates to the fields of deep learning, natural language processing and the like, in particular to a chapter-level event extraction method and device based on a multi-granularity entity differential graph.
Background
An event is a change in an event or state triggered by a verb and consisting of one or more actions engaged by one or more characters at a particular point in time or time, within a particular geographic area. Event Extraction (EE) aims at extracting structured information (such as time, place, people and the like) from an unstructured Natural Language text, is an important research task in the field of Natural Language Processing (NLP), and has wide application in the fields of intelligent question answering, information retrieval, automatic summarization, recommendation and the like.
With the development of machine learning, particularly deep learning, the sentence-level event extraction method can effectively capture the relationship and semantic information between event arguments, and greatly improves the event extraction effect. However, events in most language scenarios are often described by multiple sentences, and therefore, in order to further expand the application scenarios of event extraction, more and more researchers are beginning to focus on chapter-level event extraction. Most of the existing event extraction is mainly researched on a data set containing trigger words, and events with unobvious trigger words or without trigger words exist in a chapter, so that the common sentence-level event extraction method based on the trigger words cannot play a good role in chapter-level event extraction. In order to better verify the effect of the conventional method on a chapter-level event extraction data set lacking trigger words, some researchers redesign a chapter-level event extraction task and construct a chapter-level event extraction data set ChFinAnn not containing trigger words.
Studies based on the chfinnann dataset generally divide chapter-level event extraction into three subtasks: candidate entity extraction, event type detection and argument identification. Wherein, the candidate entity extraction is to extract entities related to the events from the texts; the event type detection is to judge which event types exist in the text; argument identification is the identification of an argument belonging to an event among candidate entities. Obviously, the candidate entity extraction is the first subtask extracted as a chapter-level event, affecting the effect on the subsequent two subtasks. The entities are distributed among multiple sentences of the chapters, and therefore, the models need to fully understand the context information before the context information can be accurately extracted. However, in the existing work, in the process of entity extraction, extraction is still performed in a sentence-by-sentence manner, semantic information of cross-sentences is not fully considered, which is not favorable for extracting entities dispersed in a plurality of sentences, and can reduce the accuracy of the entities, thereby causing problems of entity missing, label error, boundary error and the like, which can cause incompleteness and inaccuracy of arguments of events, and affect the final event extraction effect.
Disclosure of Invention
Aiming at the problems, the invention provides a chapter-level event extraction method and device based on a multi-granularity entity differential map. The method and the device respectively extract entities by using context information based on sentences and paragraphs, and fuse entity sets of two granularities based on a multi-granularity entity selection strategy. Then, combining the sentences and the screened candidate entities, constructing a heterogeneous Graph fused with the multi-granularity entities, obtaining vectorized representation of the entities and the sentences with chapter-level context perception by using a Graph volume Network (GCN), performing multi-label classification of event types and event arguments, and realizing event detection and argument identification.
In order to achieve the purpose, the invention adopts the following technical scheme
In a first aspect, the present invention provides a chapter-level event extraction device based on a multi-granularity entity heterogeneous graph, including an encoder module, a sentence-level entity extraction module, a paragraph-level entity extraction module, a multi-granularity fusion module, a heterogeneous graph construction module, an event detection module, and an argument identification module;
the encoder module comprises a sentence-level encoder and a paragraph-level encoder, which are respectively used for encoding texts with sentence granularity and paragraph granularity in chapters to obtain semantic vectorization representation of each word or word in the sentences and the paragraphs;
the sentence-level entity extraction module extracts entities from texts with sentence granularity;
the paragraph level entity extraction module extracts entities from the text of paragraph granularity;
the multi-granularity fusion module fuses entities from sentence and paragraph granularities according to rules;
the heterogeneous graph building module connects sentences and entities through defined rules, generates information interaction among cross-sentences based on a graph convolution network, and obtains vectorized representation of the sentences and the entities sensed by full text;
the event detection module is used for carrying out a plurality of secondary classifications based on vectorization representation of sentences sensed by full text so as to judge whether a certain event is triggered;
the argument identification module identifies arguments in the set of candidate entities in a path-expanded manner.
In a second aspect, the present invention provides a chapter-level event extraction method based on a multi-granularity entity heterogeneous composition, including the following steps:
step 1: the method comprises the steps that sentences and paragraph texts in chapters are respectively encoded through two independent encoders in an encoder module, and semantic vectorization representation of each word or phrase in the sentences and the paragraphs is obtained;
and 2, step: utilizing the sentence-level entity extraction module to extract entities in the sentences based on the semantic vectorization representation of the sentences obtained in the step 1;
and step 3: utilizing the paragraph level entity extraction module to extract entities in the paragraphs based on the semantic vectorization representation of the paragraphs obtained in step 1;
and 4, step 4: fusing the entities with the two granularities from the step 2 and the step 3 by using the multi-granularity fusion module, and improving the precision of the entities;
and 5: establishing connection between sentences and entity nodes by using the heteromorphic graph construction module based on defined rules, and establishing cross-sentence information interaction based on a graph convolution network to obtain the vectorized representation of sentences and entities with chapter-level context sensing;
step 6: classifying for a plurality of times by using the event detection module based on the vectorized representation of the sentence with chapter-level context awareness obtained in the step 5 to detect the event type triggered in the text;
and 7: and identifying the argument of each event by using the argument identification module, obtaining the argument of each event, obtaining a final structured event, and finishing chapter-level event extraction.
Further, the specific operation of step 1 is to use a transform as an encoder for encoding, and the calculation formula is as follows:
{t1,t2,...,ti,...,tn}＝Transformer({w1,w2,...,wi,...,wn})
wherein n is the number of sentences in the chapters, wiRepresenting the ith token in the sentence, which is also the input to the transform encoder model,
Figure BDA0003578163420000041
representing the vectorized representation after the ith sentence is coded, wherein S represents the length of the input sequence and is 128; d represents the dimension of the hidden layer, 1024.
Coding each sentence of the chapters one by one according to the formula;
at the same time, the sentence s is followed by means of dynamic sliding window1Beginning to build a paragraph
Figure BDA0003578163420000042
Figure BDA0003578163420000043
Token sequence in this paragraph
Figure BDA0003578163420000044
The length of the elastic element is less than or equal to a preset maximum length M, wherein M is
Figure BDA0003578163420000045
The number of middle tokens;the window is then slid backwards to generate the next paragraph, thus dividing an article into K paragraphs P ═ P1,...,pK}; then, the words or phrases in each paragraph are encoded using the RoBERTa model, a specific formula is as follows:
inputi＝[CLS]+Ti+[SEP]
x＝RoBERTa_wwm_ext(inputi)
wherein, TiFor the text in the ith paragraph in the chapter, the length is 510, [ CLS]Indicates the starting position, [ SEP]Denotes a separator, inputiRepresenting the input to the RoBERTa model,
Figure BDA0003578163420000046
representing a vectorized representation of paragraph text, d representing the hidden layer dimension, base version 768, large version 1024.
Further, the concrete operation of extracting the entities in the sentence in the step 2 is as follows: the vectorized representation of each sentence is subjected to multi-label classification through a full connection layer FFN, and a Viterbi algorithm in a conditional random field is used for decoding a label sequence with the maximum probability, and the calculation formula is as follows:
T＝FFN({t1,t2,...,ti,...,tn})
wherein,
Figure BDA0003578163420000051
indicating the probability, num, of each word or phrase belonging to each categorytagThe number of classified labels.
During training, the CRF is adopted to maximize the logarithmic probability of a correct label as loss; in decoding, a Viterbi algorithm is adopted to decode the tag sequence with the maximum probability, and the specific formula is as follows:
Figure BDA0003578163420000052
Zsent＝Viterbi(T)
wherein,
Figure BDA0003578163420000053
finger input sequence siThe gold tag sequence of (1), PiPointing to an input sequence siThe predicted tag sequence with the highest probability,
Figure BDA0003578163420000054
representing the resulting tag sequence;
further, the specific operation of extracting the entities in the paragraphs in step 3 is as follows: the context is better understood by using a BilSTM auxiliary model, then the vectorized representation of each paragraph is subjected to multi-label classification through a full connection layer FFN, and finally the CRF is used for identifying the entity. The specific formula is as follows:
g＝BiLSTM(x)
Score＝FFN(g)
wherein,
Figure BDA0003578163420000055
for vectorized representation of x after passing through T layers of BiLSTM,
Figure BDA0003578163420000056
Figure BDA0003578163420000057
denotes the fraction of g obtained by the FFN of the full connection layer
During training, sequence labeling takes the CRF to maximize the log probability of the correct label as a loss. In decoding, the sequence of labels having the largest probability is decoded using the Viterbi algorithm. The specific formula is as follows:
Figure BDA0003578163420000058
Zpara＝Viterbi(Score)
wherein,
Figure BDA0003578163420000061
finger input sequence piThe sequence of the gold tag of (a),
Figure BDA0003578163420000062
the paragraph entity is represented to extract the finally obtained tag sequence.
Further, the rule for fusing the entities from the two granularities of sentences and paragraphs in step 4 is as follows: selecting entities that coexist in a set of entities of two granularities. The existence of a certain entity in both granularity entity sets indicates that the credibility of the entity is higher;
selecting entities that exist only in paragraph level entity sets. The paragraph level entity extraction model takes paragraphs containing a plurality of sentences as input for extraction, and more context information can bring more accurate results;
selecting entity which exists only in a certain sentence in sentence level and also exists in other sentences in the paragraph where the sentence is. An entity exists in a plurality of sentences of a paragraph, and the existence of the entity is indicated to be reasonable.
Further, the specific operation of step 5 is that the heterogeneous graph is composed of entity nodes and sentence nodes. For an entity node, since one entity node e may contain multiple tokens, an average pooling policy is used to obtain an initialized representation of the entity node; similarly, for a sentence node, the maximum pooling strategy is applied to the token in the sentence, and the position code of the sentence is added to obtain the initialized representation of the sentence node. The specific formula is as follows:
he＝MeanPooling({ti}i∈e)
Figure BDA0003578163420000063
wherein h iseFor the initial presentation of the entity node e,
Figure BDA0003578163420000064
as sentence node siIs indicated.
When constructing edges, four types of edges are constructed using the following rules: connecting edges of all sentence nodes; linking edges between the sentence nodes and the entity nodes in the sentence; connecting edges of all entity nodes in the same sentence; fourthly, the same entity in different sentences is mentioned and connected.
After the heteromorphic image is constructed, information is transferred through the GCN of the L layer. For a node u at level i, its representation is updated by the following formula:
Figure BDA0003578163420000071
wherein, WlIs a parameter that can be learned, σ is the activation function,
Figure BDA0003578163420000072
neighbor nodes representing node u, cu,vIn order to be a normalization constant, the method comprises the following steps of,
Figure BDA0003578163420000073
is a vector representation of node u in layer l,
Figure BDA0003578163420000074
is a vector representation of a node u in the layer l +1, and v represents
Figure BDA0003578163420000075
One node in the set;
then by concatenating the node u representations of each layer, by a learnable parameter WaA linear transformation is performed to obtain the final representation of node u:
Figure BDA0003578163420000076
finally, the same entity-mention embeddings are merged into a single embeddings, again using the max-pooling strategy: e.g. of the typei＝Mean({hj}j∈Mention(i)) Wherein indication (i) denotes the set of i-th entity mentions, hjRepresenting the j-th entity in the setReference to vector representation, eiThe table is the ith entity final vector representation;
after this stage, an entity representation with chapter-level context awareness is obtained
Figure BDA0003578163420000077
Figure BDA0003578163420000078
And representation of sentences
Figure BDA0003578163420000079
Wherein N issIs the number of sentences, NeNumber mentioned for different entities, dmRepresenting the dimensions of the hidden layer.
Further, the specific operation of step 6 is: a chapter may contain multiple events and thus the task may be viewed as a multi-label classification task. And further mining the perception degree of the sentence to the event by utilizing a multi-head attention mechanism based on the feature matrix S of the sentence obtained in the last step. After multi-head attention calculation is performed on S, the calculation result is input into C classifiers for secondary classification to judge whether each event type r is triggered. Wherein C is the number of event types. The specific formula is as follows:
A＝MultiHead(Q,S,S)
rt＝FFN(A)
R＝Concat(r1,...,rC)
wherein,
Figure BDA00035781634200000710
are trainable derived parameters.
Figure BDA00035781634200000711
A score for each event type.
During training, R and gold label are combined
Figure BDA0003578163420000081
Calculating cross entropy loss, in particularThe formula is as follows:
Figure BDA0003578163420000082
further, the specific operation of step 7 is: there may be multiple events in a chapter, each event may have multiple event arguments, and the same argument may be the role of a different event. Thus, the present invention uses a path extension approach to decode event arguments. Meanwhile, in order to model the dependency between events, a Tracker module is further introduced to improve the performance.
Specifically, for each event type, the order of event roles is predefined. Then, the expansion is performed step by step starting from the first role, wherein each expanded node is either an entity or an empty node. At each step of the expansion, it is formalized as a two-class problem, i.e., it is determined whether each entity is to be expanded. Since there may be multiple eligible entities in the event role, multiple branches may be generated as the node expands. Thus, each path can be viewed as a set of arguments to an event. For an event argument path composed of entity sequences, entities in the path are spliced to obtain a representation U of the pathi＝[Ei1,...,Eie]In which Ei1And
Figure BDA0003578163420000083
each represents a vector representation of an entity in the path. It is then encoded using LSTM and converted to vector G plus embedding of the event typeiAnd then stored in a Global Memory (Global Memory). In identifying the argument of the J-th role of other events, a new representation of each entity is obtained by adding role name embedding to the entity
Figure BDA0003578163420000084
Wherein RoleJRefers to the embedding of the jth role name. Subsequently, entities are embedded
Figure BDA0003578163420000085
Sentence characteristics S, current path
Figure BDA0003578163420000086
After being spliced with a global memory G, the data are input into a Transformer to obtain a new entity characteristic matrix
Figure BDA0003578163420000087
Wherein epsilon represents the number of entities, and the specific formula is as follows:
Figure BDA0003578163420000088
wherein
Figure BDA0003578163420000089
And
Figure BDA00035781634200000810
are respectively as
Figure BDA00035781634200000814
And G is a new representation obtained after the Transformer. Considering path expansion as a plurality of two-classification problems, namely, pairs
Figure BDA00035781634200000812
Each entity in
Figure BDA00035781634200000813
Classifying to judge whether path expansion is carried out or not, and adopting the following formula as a loss function:
Figure BDA0003578163420000091
wherein, NDIs a collection of nodes in the path,
Figure BDA0003578163420000092
refers to a gold label.
In a third aspect, the present invention provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, where the processor implements the foregoing chapter-level event extraction method based on multi-granular entity difference maps when executing the program.
In a fourth aspect, the present invention provides a non-transitory computer-readable storage medium, on which a computer program is stored, where the computer program, when executed by a processor, implements the chapter-level event extraction method based on the multi-granularity entity difference map.
Compared with the prior art, the invention has the following advantages:
(1) the invention provides a chapter-level event extraction method and device based on a multi-granularity entity differential graph, wherein the multi-granularity refers to sentence granularity and paragraph granularity. Although the entity can be extracted only by using the text information of sentence granularity, the accuracy of the extracted entity is insufficient due to lack of enough context information. The invention not only extracts the sentence level entity, but also extracts the paragraph level entity, thereby effectively utilizing the context information and improving the accuracy of the entity.
(2) The sentence-level entity set and the paragraph-level entity set respectively have the problems of entity loss, boundary error, redundancy and the like, so the invention provides a fusion strategy for screening the entity sets with two granularities, which can further improve the quality of entity extraction and further improve the event extraction effect;
(3) in order to generate remote information interaction, the screened candidate entities and sentences are combined and connected through some association rules, a heterogeneous graph fused with multi-granularity entities is constructed, information exchange is performed among nodes through a graph convolution network, and accordingly vector representation of entities and sentences with full-text perception is generated.
Drawings
FIG. 1 is a diagram of a model framework of the present invention;
FIG. 2 is a sample of data used by the present invention;
FIG. 3 is a diagram of a sentence-level entity extraction module according to the present invention;
FIG. 4 is a block diagram of a paragraph level entity extraction module according to the present invention;
FIG. 5 is an algorithm of the multi-granularity fusion module of the present invention;
FIG. 6 is pseudo code of an entity fusion algorithm of the multi-granularity fusion module of the present invention;
FIG. 7 is a block diagram of the heterogeneous graph structure modeling of the present invention;
FIG. 8 is a block diagram of an event detection module of the present invention;
FIG. 9 is a block diagram of an argument identification module of the present invention;
fig. 10 is a schematic structural diagram of a chapter-level event extraction device based on a multi-granularity entity differential map according to an embodiment of the present invention;
fig. 11 is a schematic view of an overall structure of an electronic device according to an embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to the following examples and the accompanying drawings, wherein the examples are provided for the purpose of illustration only and are not intended to limit the scope of the present invention.
Example 1
Fig. 1 is an overall architecture diagram of a chapter-level event extraction method based on a multi-granularity entity heterogeneous diagram according to an embodiment of the present invention, which mainly includes an encoder module, a sentence-level entity extraction module, a paragraph-level entity extraction module, a multi-granularity fusion module, a heterogeneous diagram construction module, an event detection module, and an argument identification module.
The specific contents of each module are as follows:
the encoder module comprises a sentence-level encoder and a paragraph-level encoder, and the sentence-level encoder and the paragraph-level encoder are used for respectively encoding sentences and paragraphs in the article to obtain semantic vectorization representation of each word or each word in the sentences and the paragraphs.
And the sentence-level entity extraction module is used for carrying out multi-label classification on the vectorized representation of each token in the sentence through a full connection layer and identifying the entity by using the CRF.
And the paragraph level entity extraction module is used for further extracting features by passing the vectorized representation of each token in the paragraph through the BilSTM, then performing multi-label classification by using the full connection layer, and identifying the entity by using the CRF.
And the multi-granularity fusion module is used for screening the entity set obtained by the sentence-level entity extraction module and the entity set obtained by the paragraph-level entity extraction module according to a certain rule so as to improve the entity precision.
The heterogeneous graph construction module is used for connecting the sentences and the entity nodes according to a certain rule to construct a heterogeneous graph, and transmitting information through the GCN to obtain entity and sentence representations with chapter-level context sensing.
And the event detection module is used for realizing secondary classification of event detection by expressing the context-aware sentences with chapter levels through a plurality of full connection layers.
And the argument identification module is used for identifying the expression of the discourse-level context-aware entity in a path expansion mode, and classifying each entity in two categories to judge whether the entity is expanded or not at each expansion step. Meanwhile, after one type of event is processed, arguments of each event are stored in the global memory in sequence, and when other types of events are processed, information in the global memory is taken out and utilized.
Example 2
Fig. 2 is an example of a chinese financial chapter-level event extraction dataset chfinnann, which originates from a real financial event, where the triggering event is a stock freeze and the argument of the event is distributed among multiple sentences of the context.
1. Firstly, the encoder module is used for encoding the chapters sentence by sentence to obtain the vector representation of each word or phrase in the sentence. The invention adopts a Transformer as an encoder to encode the data, and the calculation formula is as follows:
{t1,t2,...,ti,...,tn}＝Transformer({w1,w2,...,wi,...,wn})
where n is the number of sentences in the chapter, wiRepresenting the ith token in the sentence, which is also the input to the Transformer encoder model,
Figure BDA0003578163420000121
representing the encoded vector representation of the ith sentence. Where S represents the input sequence length 128; d represents the dimension 1024 of the hidden layer.
At the same time, the sentence s is followed by means of dynamic sliding window1Beginning to build a paragraph
Figure BDA0003578163420000122
Figure BDA0003578163420000123
Token sequence in this paragraph
Figure BDA0003578163420000124
The length of the elastic element is less than or equal to a preset maximum length M, wherein M is
Figure BDA0003578163420000125
Number of middle tokens. The window is then slid backwards to generate the next paragraph, thus dividing an article into K paragraphs P ═ P1,...,pK}. Then, the word or phrase in each paragraph is encoded using the training language model RoBERTa, with the following specific formula:
inputi＝[CLS]+Ti+[SEP]
x＝RoBERTa_wwm_ext(input i)
wherein, TiIs the text in the ith paragraph in the chapter and has a length of 510. [ CLS]Indicates the starting position, [ SEP]Representing a separator. inputiRepresenting the input to the RoBERTa model,
Figure BDA0003578163420000126
represents the vector representation of paragraph text, d represents the hidden layer dimension, base version is 768, and large version is 1024.
2. Entities in sentences are extracted by using a sentence-level entity extraction module, and fig. 3 is a structural diagram of the module. After each sentence is coded one by one, multi-label classification is carried out on the vectorization representation of each sentence through a full connection layer FFN, and a Viterbi algorithm in a Conditional Random Field (CRF) is used for decoding a label sequence with the maximum probability to obtain an entity in the sentence. The calculation formula is as follows:
T＝FFN({t1,t2,...,ti,...,tn})
Zsent＝Viterbi(T)
wherein,
Figure BDA0003578163420000127
indicating the probability, num, of each word or phrase belonging to each categorytagIn order to determine the number of labels to be classified,
Figure BDA0003578163420000128
representing the resulting tag sequence;
3. the entity in the paragraph is extracted by using a paragraph-level entity extraction module, fig. 4 is a structural diagram of the module, a BiLSTM auxiliary model is used to better understand the context, then the vectorization representation of each paragraph is subjected to multi-label classification through a full connection layer FFN, and finally, a Viterbi algorithm in the CRF is used to decode the label sequence with the maximum probability to obtain the entity in the paragraph. The calculation formula is as follows:
g＝BiLSTM(x)
Score＝FFN(g)
Zpara＝Viterbi(Score)
wherein,
Figure BDA0003578163420000131
is the vectorized representation of x after passing through the T layer BilSTM,
Figure BDA0003578163420000132
Figure BDA0003578163420000133
represents the fraction of u found through the full connection layer FFN,
Figure BDA0003578163420000134
indicates the final resultThe tag sequence of (1).
4. The entity from two granularities is fused by a multi-granularity fusion module to improve the precision of the entity, fig. 5 is a structure diagram of the module, and fig. 6 is a pseudo code of the module fusion method. The fusion rule is as follows: selecting entities with two granularities which coexist in a set; selecting entities that exist only in the paragraph level entity set; selecting entity which exists only in a certain sentence in sentence level and also exists in other sentences in the paragraph where the sentence is.
5. The heterogeneous graph building module is used for establishing a relation between sentences and entity nodes, and fig. 7 is a structure diagram of the module. The abnormal graph is composed of entity nodes and sentence nodes. For an entity node, since one entity node e may contain multiple tokens, an average pooling policy is used to obtain an initialized representation of the entity node; similarly, for a sentence node, the maximum pooling strategy is applied to the token in the sentence, and the position of the sentence is encoded to obtain the initialized representation of the sentence node. The specific formula is as follows:
he＝MeanPooling({ti}i∈e)
Figure BDA0003578163420000135
wherein h iseFor an initial representation of the entity node e,
Figure BDA0003578163420000136
as sentence node siIs indicated.
When constructing edges, four types of edges are constructed using the following rules: connecting edges of all sentence nodes; linking edges between the sentence nodes and the entity nodes in the sentence; connecting edges of all entity nodes in the same sentence; fourthly, the same entity in different sentences is mentioned and connected.
After the heteromorphic image is constructed, information is transferred through the GCN of the L layer. For a node u at level i, its representation is updated by the following formula:
Figure BDA0003578163420000141
wherein, WlIs a parameter that can be learned, σ is the activation function,
Figure BDA0003578163420000142
neighbor nodes representing node u, cu,vIn order to be a normalization constant, the method comprises the following steps of,
Figure BDA0003578163420000143
is a vector representation of node u in layer l,
Figure BDA0003578163420000144
is a vector representation of node u in layer l +1, and v represents
Figure BDA0003578163420000145
One node in the set.
Then by concatenating the node u representations of each layer, by a learnable parameter WaA linear transformation is performed to obtain the final representation of node u:
Figure BDA0003578163420000146
finally, the same entity mention embeddings are merged into a single embeddings, again using the max pooling strategy: e.g. of the typei＝Mean({hj}j∈Mention(i)) Wherein indication (i) denotes the set of i-th entity mentions, hjVector representation of a reference representing the jth entity in the set, eiThe table is the ith entity final vector representation. After this stage, an entity representation with discourse-level context awareness is obtained
Figure BDA0003578163420000147
And vectorized representation of sentences
Figure BDA0003578163420000148
Wherein N issIs the number of sentences, NeNumber mentioned for different entities, dmRepresenting the dimensions of the hidden layer.
6. An event detection module is used for detecting an event in the text, and fig. 8 is a block diagram of the module. A chapter may contain multiple events and thus the task may be viewed as a multi-label classification task. And obtaining a feature matrix S of the sentence after the last step, and further mining the perception degree of the sentence on the event by using a multi-head attention mechanism. After multi-head attention calculation is performed on S, the calculation result is input into C classifiers for secondary classification to judge whether each event type r is triggered. Wherein C is the number of event types. The specific formula is as follows:
A＝MultiHead(Q,S,S)
rt＝FFN(A)
R＝Concat(r1,...,rC)
wherein,
Figure BDA0003578163420000151
are trainable derived parameters.
Figure BDA0003578163420000152
A score for each event type.
7. The argument of each event is identified by an argument identification module, of which block diagram fig. 9 is a diagram. For each event type, the order of event roles is predefined. Then, the expansion is performed step by step starting from the first role, wherein each expanded node is either an entity or an empty node. At each step of the expansion, it is formalized as a two-class problem, i.e., it is determined whether each entity is to be expanded. Since there may be multiple eligible entities in the event role, multiple branches may be generated as the node expands. Thus, each path can be viewed as a set of arguments to an event. For an event argument path composed of entity sequences, entities in the path are spliced to obtain a representation of the path
Figure BDA0003578163420000153
Wherein Ei1And
Figure BDA0003578163420000154
each represents a vector representation of an entity in the path. It is then encoded using LSTM and converted to vector G plus embedding of the event typeiAnd then stored in a Global Memory (Global Memory). In identifying the argument of the J-th role of other events, a new representation of each entity is obtained by adding role name embedding to the entity
Figure BDA0003578163420000155
Wherein RoleJRefers to the embedding of the jth role name. Subsequently, entities are embedded
Figure BDA0003578163420000156
Sentence feature S, Current Path
Figure BDA0003578163420000157
After being spliced with a global memory G, the data are input into a Transformer to obtain a new entity characteristic matrix
Figure BDA0003578163420000158
The specific formula is as follows:
Figure BDA0003578163420000159
wherein
Figure BDA00035781634200001510
And
Figure BDA00035781634200001511
are respectively as
Figure BDA00035781634200001512
And G is a new representation obtained after the Transformer. Path expansion is treated as a number of two-class problems, i.e.
Figure BDA00035781634200001513
Each entity in
Figure BDA00035781634200001514
Classification is performed to determine whether or not to perform path expansion.
Example 3
Fig. 10 is a schematic structural diagram of a chapter-level event extraction device based on a multi-granularity entity differential map according to an embodiment of the present invention. The chapter-level event extraction device comprises: the system comprises an encoder module, an entity extraction module, a multi-granularity fusion module, a heterogeneous graph construction module, an event detection module and an argument identification module.
The encoder module consists of two parts, namely a sentence encoder and a paragraph encoder, which are respectively used for encoding the sentence and paragraph granularity text to obtain vectorization representation of each character or word in the sentence and the paragraph;
the entity extraction module comprises a sentence-level entity extraction module and a paragraph-level entity extraction module which are respectively used for extracting entities from sentences and paragraphs;
the multi-granularity fusion module fuses entities from sentences and paragraphs to improve entity precision;
the heterogeneous graph construction module is used for connecting sentences and entity nodes according to a certain rule to construct a heterogeneous graph, and transmitting information through the GCN to obtain entity and sentence representations with chapter-level context sensing.
And the event detection module is used for realizing secondary classification of event detection by expressing the context-aware sentences with chapter levels through a plurality of full connection layers.
And the argument identification module is used for identifying the expression of the discourse-level context-aware entity in a path expansion mode, and classifying each entity in two categories to judge whether the entity is expanded or not at each expansion step. Meanwhile, after one type of event is processed, arguments of each event are stored in the global memory in sequence, and when other types of events are processed, information in the global memory is taken out and utilized.
The embodiment of the present invention provides a chapter-level event extraction device based on a multi-granularity entity differential map, which specifically executes the flow of the chapter-level event extraction method based on the multi-granularity entity differential map, and please refer to the content of the chapter-level event extraction method based on the multi-granularity entity differential map in detail, which is not described herein again.
This embodiment provides an electronic device, and fig. 11 is a schematic diagram of an overall structure of the electronic device according to the embodiment of the present invention, where the electronic device includes: a processor, a memory, a communication bus, and a communication interface; the processor, the communication interface and the memory are communicated with each other through a communication bus. The memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the methods provided by the method embodiments, for example, the method includes: the method comprises the steps that sentences and paragraphs are coded through a coder, and semantic representations of the sentences and paragraphs are output; obtaining the classification probability of each word in a sentence or paragraph by using a full-connection network; obtaining a vector representation of sentences and entities with full-text perception based on a graph convolution network; obtaining the probability of the type of the trigger event by utilizing a self-attention mechanism and a fully-connected network; and identifying arguments based on the path expansion mode.
In addition, the logic instructions in the memory may be implemented in the form of software functional units, and may be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the embodiments of the present invention may be embodied in the form of a software product, which is stored in a storage medium and includes instructions for causing a computer device to perform all or part of the steps of the methods according to the embodiments of the present invention. And the aforementioned storage medium includes: a U disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), and other various media capable of storing program codes.
The present embodiments provide a non-transitory computer-readable storage medium storing computer instructions that cause a computer to perform the methods provided by the above method embodiments, for example, including: the method comprises the steps that sentences and paragraphs are coded through a coder, and semantic representations of the sentences and paragraphs are output; obtaining the classification probability of each word in a sentence or paragraph by using a full-connection network; obtaining a vector representation of sentences and entities with full-text perception based on a graph convolution network; obtaining the probability of the type of the trigger event by utilizing a self-attention mechanism and a fully-connected network; and identifying arguments based on the path expansion mode.
The above-described embodiments of the apparatus are merely illustrative, and the units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.
The above embodiments are only for illustrating the technical solutions of the present invention and are not limited thereto, and any modification or equivalent replacement without departing from the spirit and scope of the present invention should be covered within the technical solutions of the present invention.

Claims (10)

1. A chapter-level event extraction device based on a multi-granularity entity differential composition is characterized in that: the system comprises an encoder module, a sentence-level entity extraction module, a paragraph-level entity extraction module, a multi-granularity fusion module, an isomerous graph construction module, an event detection module and an argument identification module;
the encoder module comprises a sentence-level encoder and a paragraph-level encoder, which are respectively used for encoding texts with sentence granularity and paragraph granularity in chapters to obtain semantic vectorization representation of each word or word in the sentences and the paragraphs;
the sentence-level entity extraction module extracts entities from texts with sentence granularity;
the paragraph level entity extraction module extracts entities from the text of paragraph granularity;
the multi-granularity fusion module fuses entities from sentence and paragraph granularities according to rules;
the heterogeneous graph building module connects sentences and entities through defined rules, generates information interaction among cross-sentences based on a graph convolution network, and obtains full-text perception of sentences and vectorization representation of the entities;
the event detection module is used for carrying out a plurality of secondary classifications based on vectorized representation of sentences with full-text perception so as to judge whether a certain event is triggered;
the argument identification module identifies arguments in the set of candidate entities in a path-expanded manner.
2. A chapter-level event extraction method based on a multi-granularity entity differential composition is characterized by comprising the following steps: the method comprises the following steps:
step 1: the method comprises the steps that sentences and paragraph texts in chapters are respectively encoded through two independent encoders in an encoder module, and semantic vectorization representation of each word or phrase in the sentences and the paragraphs is obtained;
step 2: utilizing the sentence-level entity extraction module to extract entities in the sentences based on the semantic vectorization representation of the sentences obtained in the step 1;
and step 3: utilizing the paragraph level entity extraction module to extract entities in the paragraphs based on the semantic vectorization representation of the paragraphs obtained in step 1;
and 4, step 4: fusing the entities with the two granularities from the step 2 and the step 3 by utilizing the multi-granularity fusion module;
and 5: establishing connection between sentences and entities by using the heteromorphic graph construction module based on defined rules, and establishing cross-sentence information interaction based on a graph convolution network to obtain the vectorized representation of the sentences and the entities with chapter-level context sensing;
step 6: classifying for a plurality of times by using the event detection module based on the vectorized representation of the sentence with chapter-level context awareness obtained in the step 5 to detect the event type triggered in the text;
and 7: and identifying the argument of each event by using the argument identification module, obtaining the argument of each event, obtaining a final structured event, and finishing chapter-level event extraction.
3. The method for extracting discourse-level events based on the multi-granularity entity heterogeneous composition as claimed in claim 2, wherein: the specific operation of step 1 is to use a transform as an encoder for encoding, and the calculation formula is as follows:
{t1，t2，...，ti，...，tn}＝Transformer({w1，w2，...，wi，...，wn})
wherein n is the number of sentences in the chapter, wiRepresenting the ith token in the sentence, which is also the input to the transform encoder model,
Figure FDA0003578163410000021
represents the encoded vectorized representation of the ith sentence, where S represents the length of the input sequence, which is 128; d represents the dimension of the hidden layer, and is 1024; coding each sentence of the chapters one by one according to the formula;
at the same time, the sentence s is followed by means of dynamic sliding window1Beginning to build a paragraph
Figure FDA0003578163410000022
Token sequence in this paragraph
Figure FDA0003578163410000023
The length of the elastic element is less than or equal to a preset maximum length M, wherein M is
Figure FDA0003578163410000024
The number of tokens and then sliding the window backwards to generate the next paragraph divides an article into K paragraphs P ═ { P }1，...，pKAre then, causeThe words or phrases in each paragraph are encoded using the RoBERTA training language model, the specific formula is as follows:
inputi＝[CLS]+Ti+[SEP]
x＝RoBERTa_wwm_ext(inputi)
wherein, TiFor the text in the ith paragraph in the chapter, the length is 510, [ CLS]Indicates the starting position, [ SEP]Denotes a separator, inputiRepresenting the input to the RoBERTa model,
Figure FDA0003578163410000031
representing a vectorized representation of paragraph text, d representing the hidden layer dimension, base version 768, large version 1024.
4. The method for extracting discourse-level events based on the multi-granularity entity heterogeneous composition as claimed in claim 2, wherein: the specific operation of extracting entities in the sentences in the step 2 is to perform multi-label classification on the vectorized representation of each sentence through a full connection layer FFN, and decode the label sequence with the maximum probability by using a Viterbi algorithm in a conditional random field, wherein the calculation formula is as follows:
T＝FFN({t1，t2，...，ti，...，tn})
Zsent＝Viterbi(T)
wherein,
Figure FDA0003578163410000032
indicating the probability, num, of each word or phrase belonging to each categorytagIn order to determine the number of labels to be classified,
Figure FDA0003578163410000033
representing a tag sequence finally obtained by sentence entity extraction;
the specific operation of extracting the entities in the paragraphs in the step 3 is that a BilSTM auxiliary model is used for better understanding the context, and then the vectorization representation of each paragraph is subjected to multi-label classification through a full connection layer FFN to obtain the probability sequence of each token belonging to each label; finally, decoding the label sequence with the maximum probability by using a Viterbi algorithm in the conditional random field; the specific calculation formula is as follows:
g＝BiLSTM(x)
Score＝FFN(g)
Zpara＝Viterbi(Score)
wherein,
Figure FDA0003578163410000034
is the vectorized representation of x after passing through the T layer BilSTM,
Figure FDA0003578163410000035
Figure FDA0003578163410000036
represents the fraction of g found through the full connection layer FFN,
Figure FDA0003578163410000037
the paragraph entity is represented to extract the finally obtained tag sequence.
5. The method for extracting discourse-level events based on the multi-granularity entity heterogeneous composition as claimed in claim 2, wherein: the rule for fusing the entities from the two granularities of sentences and paragraphs in the step 4 is as follows: selecting entities with two granularities which coexist in a set; selecting entities that exist only in the paragraph level entity set; selecting entity which exists only in a certain sentence in sentence level and also exists in other sentences in the paragraph where the sentence is.
6. The method for extracting discourse-level events based on the multi-granularity entity heterogeneous composition as claimed in claim 2, wherein: the concrete operation of the step 5 is that the heteromorphic graph is composed of entity nodes and sentence nodes; for an entity node, since one entity node e may contain multiple tokens, an average pooling policy is used to obtain an initialized representation of the entity node; similarly, for a sentence node, using a maximum pooling strategy for token in the sentence and adding the position code of the sentence to obtain the initialized representation of the sentence node; the specific formula is as follows:
he＝MeanPooling({ti}i∈e)
Figure FDA0003578163410000041
wherein h iseFor an initial representation of the entity node e,
Figure FDA0003578163410000042
as nodes s of sentencesiAn initialization representation of (a);
when constructing edges, four types of edges are constructed using the following rules: connecting edges of all sentence nodes; linking edges between the sentence nodes and the entity nodes in the sentence; connecting all the entity nodes in the same sentence; fourthly, the same entity in different sentences is mentioned and connected;
after the abnormal graph is constructed, information is transmitted through the GCN of the L layer, and the expression of the node u of the L-th layer is updated through the following formula:
Figure FDA0003578163410000043
wherein, WlIs a parameter that can be learned, σ is the activation function,
Figure FDA0003578163410000044
a neighbor node representing node u, cu，vIn order to be a normalization constant, the method comprises the following steps of,
Figure FDA0003578163410000045
is a vector representation of node u in layer l,
Figure FDA0003578163410000046
is a vector representation of a node u in the layer l +1, and v represents
Figure FDA0003578163410000047
One node in the set;
then by concatenating the node u representations of each layer, by a learnable parameter WaA linear transformation is performed to obtain the final representation of node u:
Figure FDA0003578163410000051
finally, the same entity mention embeddings are merged into a single embeddings, again using the max pooling strategy: e.g. of the typei＝MeanPooling({hj}j∈Mention(i)) Wherein indication (i) denotes the set of i-th entity mentions, hjVector representation of a mention of the jth entity in the set, eiThe table is the ith entity final vector representation;
after this stage, a vectorized representation of the entity with chapter-level context awareness is obtained
Figure FDA0003578163410000052
And vectorized representation of sentences
Figure FDA0003578163410000053
Wherein N issIs the number of sentences, NeNumber of references to different entities, dmRepresenting the dimensions of the hidden layer.
7. The method for extracting discourse-level events based on the multi-granularity entity heterogeneous composition as claimed in claim 2, wherein: the specific operation of the step 6 is to further mine the perception degree of the sentence to the event by using a multi-head attention mechanism based on the sentence characteristic matrix S; after multi-head attention calculation is carried out on S, the calculation result is input into C classifiers for secondary classification to judge whether each event type r is triggered; wherein, C is the number of the event types, and the specific formula is as follows:
A＝MultiHead(Q，S，S)
rt＝FFN(A)
R＝Concat(r1，...，rC)
wherein,
Figure FDA0003578163410000054
in order to obtain the parameters in a trainable manner,
Figure FDA0003578163410000055
a score for each event type.
8. The method for extracting discourse-level events based on the multi-granularity entity heterogeneous composition as claimed in claim 2, wherein the specific operations of the step 7 are as follows:
for each event type, the sequence of event roles is predefined; then, gradually expanding from the first role, wherein each expanded node is either an entity or an empty node; at each step of the expansion, the method is formalized into a problem of two categories, namely, whether each entity needs to be expanded or not is judged; because the event role may have a plurality of entities meeting the conditions, a plurality of branches may be generated when the node is expanded; thus, each path can be viewed as a set of arguments for an event; for an event argument path composed of entity sequences, the entities in the path are spliced to obtain a representation of the path
Figure FDA0003578163410000061
Wherein Ei1And
Figure FDA0003578163410000062
vector representations each representing an entity in the path; it is then encoded using LSTM and converted to vector G plus embedding of the event typeiThen storing the data into a global memory; when argument of J-th role of other event is identifiedFor each entity, new representation of the entity is obtained by adding the role name embedding
Figure FDA0003578163410000063
Wherein RoleJThe embedding of the J-th role name is referred to; subsequently, entities are embedded
Figure FDA0003578163410000064
Sentence characteristics S, current path
Figure FDA0003578163410000065
After being spliced with a global memory G, the obtained object is input into a Transformer to obtain a new entity characteristic matrix
Figure FDA0003578163410000066
Wherein epsilon represents the number of entities, and the specific formula is as follows:
Figure FDA0003578163410000067
wherein
Figure FDA0003578163410000068
And
Figure FDA0003578163410000069
are respectively as
Figure FDA00035781634100000610
S，UjAnd G is a new representation obtained after the Transformer; considering path expansion as a plurality of two-classification problems, namely, pairs
Figure FDA00035781634100000611
Each entity in
Figure FDA00035781634100000612
Classifying to determine whether to useAnd performing path expansion.
9. An electronic device, characterized in that: the system comprises a memory, a processor and a computer program stored on the memory and running on the processor, wherein the processor implements the chapter-level event extraction method based on the multi-granularity entity difference chart according to any one of claims 2 to 8 when executing the computer program.
10. A non-transitory computer-readable storage medium, characterized in that: the computer readable storage medium has a computer program stored thereon, and when executed by a processor, the computer program implements the chapter-level event extraction method based on the multi-granularity entity difference map according to any one of claims 2 to 8.
CN202210348614.3A 2022-04-01 2022-04-01 Chapter-level event extraction method and device based on multi-granularity entity different composition Active CN114742016B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210348614.3A CN114742016B (en) 2022-04-01 2022-04-01 Chapter-level event extraction method and device based on multi-granularity entity different composition

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210348614.3A CN114742016B (en) 2022-04-01 2022-04-01 Chapter-level event extraction method and device based on multi-granularity entity different composition

Publications (2)

Publication Number Publication Date
CN114742016A true CN114742016A (en) 2022-07-12
CN114742016B CN114742016B (en) 2024-07-09

Family

ID=82278978

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210348614.3A Active CN114742016B (en) 2022-04-01 2022-04-01 Chapter-level event extraction method and device based on multi-granularity entity different composition

Country Status (1)

Country Link
CN (1) CN114742016B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116304077A (en) * 2022-12-19 2023-06-23 河海大学 Method for extracting text events of flood and drought disasters based on different patterns
CN116757159A (en) * 2023-08-15 2023-09-15 昆明理工大学 End-to-end multitasking joint chapter level event extraction method and system

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111581345A (en) * 2020-04-26 2020-08-25 上海明略人工智能(集团)有限公司 Document level event extraction method and device
CN113076483A (en) * 2021-04-27 2021-07-06 昆明理工大学 Case element heteromorphic graph-based public opinion news extraction type summarization method
CN113255321A (en) * 2021-06-10 2021-08-13 之江实验室 Financial field chapter-level event extraction method based on article entity word dependency relationship
WO2021164199A1 (en) * 2020-02-20 2021-08-26 齐鲁工业大学 Multi-granularity fusion model-based intelligent semantic chinese sentence matching method, and device
CN113591483A (en) * 2021-04-27 2021-11-02 重庆邮电大学 Document-level event argument extraction method based on sequence labeling
US20210357585A1 (en) * 2017-03-13 2021-11-18 Arizona Board Of Regents On Behalf Of The University Of Arizona Methods for extracting and assessing information from literature documents
CN113761936A (en) * 2021-08-19 2021-12-07 哈尔滨工业大学（威海） Multi-task chapter-level event extraction method based on multi-head self-attention mechanism

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210357585A1 (en) * 2017-03-13 2021-11-18 Arizona Board Of Regents On Behalf Of The University Of Arizona Methods for extracting and assessing information from literature documents
WO2021164199A1 (en) * 2020-02-20 2021-08-26 齐鲁工业大学 Multi-granularity fusion model-based intelligent semantic chinese sentence matching method, and device
CN111581345A (en) * 2020-04-26 2020-08-25 上海明略人工智能(集团)有限公司 Document level event extraction method and device
CN113076483A (en) * 2021-04-27 2021-07-06 昆明理工大学 Case element heteromorphic graph-based public opinion news extraction type summarization method
CN113591483A (en) * 2021-04-27 2021-11-02 重庆邮电大学 Document-level event argument extraction method based on sequence labeling
CN113255321A (en) * 2021-06-10 2021-08-13 之江实验室 Financial field chapter-level event extraction method based on article entity word dependency relationship
CN113761936A (en) * 2021-08-19 2021-12-07 哈尔滨工业大学（威海） Multi-task chapter-level event extraction method based on multi-head self-attention mechanism

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
张虎等: "基于多粒度实体异构图的篇章级事件抽取方法", 《计算机科学》, 31 May 2023 (2023-05-31) *
陈佳丽;洪宇;王捷;张婧丽;姚建民;: "利用门控机制融合依存与语义信息的事件检测方法", 中文信息学报, no. 08, 15 August 2020 (2020-08-15) *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116304077A (en) * 2022-12-19 2023-06-23 河海大学 Method for extracting text events of flood and drought disasters based on different patterns
CN116757159A (en) * 2023-08-15 2023-09-15 昆明理工大学 End-to-end multitasking joint chapter level event extraction method and system
CN116757159B (en) * 2023-08-15 2023-10-13 昆明理工大学 End-to-end multitasking joint chapter level event extraction method and system

Also Published As

Publication number Publication date
CN114742016B (en) 2024-07-09

Similar Documents

Publication Publication Date Title
CN108416058B (en) Bi-LSTM input information enhancement-based relation extraction method
CN112434535B (en) Element extraction method, device, equipment and storage medium based on multiple models
CN111950287B (en) Entity identification method based on text and related device
CN111723569A (en) Event extraction method and device and computer readable storage medium
CN111737974B (en) Semantic abstract representation method and device for statement
CN114742016B (en) Chapter-level event extraction method and device based on multi-granularity entity different composition
CN113886601B (en) Electronic text event extraction method, device, equipment and storage medium
CN113486178B (en) Text recognition model training method, text recognition method, device and medium
CN111145914B (en) Method and device for determining text entity of lung cancer clinical disease seed bank
CN116661805B (en) Code representation generation method and device, storage medium and electronic equipment
CN116304748B (en) Text similarity calculation method, system, equipment and medium
CN112200664A (en) Repayment prediction method based on ERNIE model and DCNN model
CN110852066B (en) Multi-language entity relation extraction method and system based on confrontation training mechanism
CN115168541A (en) Chapter event extraction method and system based on frame semantic mapping and type perception
CN116228383A (en) Risk prediction method and device, storage medium and electronic equipment
CN116150367A (en) Emotion analysis method and system based on aspects
CN113553847A (en) Method, device, system and storage medium for parsing address text
CN113239694B (en) Argument role identification method based on argument phrase
CN117670017A (en) Event-based risk identification method and device and electronic equipment
CN116432755A (en) Weight network reasoning method based on dynamic entity prototype
CN116562291A (en) Chinese nested named entity recognition method based on boundary detection
CN114330350B (en) Named entity recognition method and device, electronic equipment and storage medium
CN112487811B (en) Cascading information extraction system and method based on reinforcement learning
CN114417891A (en) Reply sentence determination method and device based on rough semantics and electronic equipment
CN114548113A (en) Event-based reference resolution system, method, terminal and storage medium

Legal Events

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