CN114943221A - Construction method of segment pointer interaction model and social sensing disaster monitoring method - Google Patents
Construction method of segment pointer interaction model and social sensing disaster monitoring method Download PDFInfo
- Publication number
- CN114943221A CN114943221A CN202210374738.9A CN202210374738A CN114943221A CN 114943221 A CN114943221 A CN 114943221A CN 202210374738 A CN202210374738 A CN 202210374738A CN 114943221 A CN114943221 A CN 114943221A
- Authority
- CN
- China
- Prior art keywords
- interval
- representation
- starting
- text
- interaction
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 230000003993 interaction Effects 0.000 title claims abstract description 146
- 238000012544 monitoring process Methods 0.000 title claims abstract description 47
- 238000000034 method Methods 0.000 title claims abstract description 34
- 238000010276 construction Methods 0.000 title claims abstract description 15
- 238000012549 training Methods 0.000 claims abstract description 100
- 238000001514 detection method Methods 0.000 claims abstract description 54
- 230000008447 perception Effects 0.000 claims abstract description 36
- 238000012545 processing Methods 0.000 claims description 35
- 238000007781 pre-processing Methods 0.000 claims description 18
- 238000001914 filtration Methods 0.000 claims description 13
- 230000011218 segmentation Effects 0.000 claims description 12
- 238000004422 calculation algorithm Methods 0.000 claims description 9
- 238000004590 computer program Methods 0.000 claims description 7
- 238000012806 monitoring device Methods 0.000 claims description 5
- 239000013598 vector Substances 0.000 description 9
- 238000010586 diagram Methods 0.000 description 8
- 230000002452 interceptive effect Effects 0.000 description 8
- 230000006870 function Effects 0.000 description 7
- 230000000694 effects Effects 0.000 description 6
- 238000002372 labelling Methods 0.000 description 4
- 230000008569 process Effects 0.000 description 4
- 238000000605 extraction Methods 0.000 description 3
- 239000000463 material Substances 0.000 description 3
- 239000011159 matrix material Substances 0.000 description 3
- 241001133760 Acoelorraphe Species 0.000 description 2
- 239000012634 fragment Substances 0.000 description 2
- 238000009432 framing Methods 0.000 description 2
- 238000007477 logistic regression Methods 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000009825 accumulation Methods 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 239000000969 carrier Substances 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 238000007499 fusion processing Methods 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Images
Classifications
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/01—Social networking
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
- G06Q50/26—Government or public services
- G06Q50/265—Personal security, identity or safety
-
- 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
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A10/00—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE at coastal zones; at river basins
- Y02A10/40—Controlling or monitoring, e.g. of flood or hurricane; Forecasting, e.g. risk assessment or mapping
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Tourism & Hospitality (AREA)
- Life Sciences & Earth Sciences (AREA)
- Marketing (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Data Mining & Analysis (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Molecular Biology (AREA)
- Primary Health Care (AREA)
- Strategic Management (AREA)
- Evolutionary Computation (AREA)
- General Business, Economics & Management (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computer Security & Cryptography (AREA)
- Development Economics (AREA)
- Educational Administration (AREA)
- Machine Translation (AREA)
Abstract
The invention discloses a construction method of a segment pointer interaction model and a social sensing disaster monitoring method, wherein the construction method of the segment pointer interaction model comprises the following steps: acquiring a training text set; inputting the training text into an entity perception coding layer to obtain semantic representation of the training text; inputting the semantic representation of the training text into a pointer network detection layer to obtain the starting interval representation and the ending interval representation of each word in the training text; inputting the starting interval representation and the ending interval representation into an interval interaction perception layer, and performing feature interaction on the starting interval representation and the ending interval representation by the interval interaction perception layer to obtain a starting category label and an ending category label after the feature interaction; and generating a corresponding starting category list and a corresponding terminating category list based on the starting category label and the terminating category label, and decoding based on the starting category list and the terminating category list to obtain the disaster event trigger word. The invention can realize effective and rapid event detection.
Description
Technical Field
The invention relates to the technical field of natural language processing, in particular to a construction method of a fragment pointer interaction model and a social sensing disaster monitoring method.
Background
The occurrence of natural disasters brings inconvenience to people, and disaster relief work needs to be carried out in time to restore normal living order. Traditional disaster situation collection work is mainly accomplished by relevant staff by hand, but limited staff hardly accomplishes the full coverage to the disaster situation incident in the short time, and partial disaster situation information needs the staff to investigate in fact and report, often can effectively master the disaster situation information after the disaster takes place for a long time, promptly, traditional disaster situation collection work consumes great manpower and materials and the timeliness is poor. In recent years, with the development of social media, social media users can conveniently send their living conditions through mobile equipment, and at the first time of disaster occurrence, the social media users play the role of a dynamic sensor as a group in which the disaster is directly contacted, and short texts sent by the social media users serve as social sensing information, and the short texts related to the disaster serve as carriers of social sensing, so that the obtaining ways of disaster conditions are greatly widened, and the information has strong real-time performance, does not need to be manually input again, and can effectively improve the disaster treatment efficiency.
The detection of the disaster event based on the short text is mainly characterized by the detection of a disaster event trigger word, wherein the disaster event trigger word refers to a word for indicating the occurrence of the disaster event, for example, "typhoon attacks the XX area of XX city positively, so that green trees of a plurality of main roads in the XX area collapse and block traffic" collapse "in the sentence is the disaster event trigger word, and the disaster event trigger word is usually detected to detect the disaster fact.
However, the prior art lacks an effective disaster event trigger word detection method based on short text of social media.
Disclosure of Invention
The invention solves the problem of how to realize a disaster event trigger word detection method based on a short text of a social media.
The invention provides a construction method of a segment pointer interaction model, wherein the segment pointer interaction model comprises an entity perception coding layer, a pointer network detection layer and an interval interaction perception layer based on a pre-training language model; the construction method of the segment pointer interaction model comprises the following steps:
acquiring a training text set, wherein the training text set is text data obtained by preprocessing social media text data;
inputting the training texts in the training text set into the entity perception coding layer to obtain semantic representation of the training texts output by the entity perception coding layer;
inputting the semantic representation of the training text into the pointer network detection layer to obtain the starting interval representation and the ending interval representation of each word in the training text predicted by the pointer network detection layer;
inputting the starting interval representation and the ending interval representation into the interval interaction perception layer, and performing feature interaction on the starting interval representation and the ending interval representation by the interval interaction perception layer to obtain a starting category label and an ending category label after feature interaction;
and generating a corresponding starting category list and a corresponding terminating category list based on the starting category label and the terminating category label, and decoding based on the starting category list and the terminating category list to obtain the disaster event trigger word.
Optionally, the pre-processing operation comprises the steps of:
extracting entity information from the social media text based on a preset word segmentation algorithm, wherein the entity information comprises an entity type of an extracted entity word and position information of the entity word in the social media text;
attaching the entity type and the position information of the entity word extracted from the social media text to be used as the training text;
the pre-training language model is a BERT model; the step of inputting the training texts in the training text set into the entity-aware coding layer to obtain semantic representations of the training texts output by the entity-aware coding layer includes an input coding step when the semantic representations are input into the entity-aware coding layer, which specifically includes:
and coding the social media texts into an upper sentence and a lower sentence, wherein one sentence corresponds to the sequence coding of the full social media texts, and the other sentence corresponds to the coding of the entity type and the position information of the entity words extracted from the social media texts.
Optionally, before the step of extracting entity information from the social media text based on the preset word segmentation algorithm, the preprocessing operation further includes the following steps:
obtaining original social media text data, and processing the original social media text data by adopting at least one of the following operations:
performing deduplication processing on the original social media text data;
filtering the original social media text data by adopting a preset keyword template, wherein keywords in the keyword template comprise disaster situation fact irrelevance texts;
and filtering the non-event sentences in the original social media text data.
Optionally, the inputting the starting interval representation and the ending interval representation into the interval interaction sensing layer, and performing feature interaction on the starting interval representation and the ending interval representation by the interval interaction sensing layer to obtain a starting category tag and an ending category tag after the feature interaction includes the following steps:
after the initial interval representation and the termination interval representation are interacted, the original semantic representation of the training text is blended to obtain a first feature;
after the first characteristic is subjected to linear processing, a second characteristic is obtained;
and outputting a starting class label or a terminating class label after the first characteristic and the second characteristic are interacted.
Optionally, the inputting the starting interval representation and the ending interval representation into the interval interaction sensing layer, performing feature interaction on the starting interval representation and the ending interval representation by the interval interaction sensing layer, and obtaining a starting category tag and an ending category tag after the feature interaction includes:
r=tanh(W D ·concat(u s ,u e )), (1)
m (1) =W (1) ·concat(h,r)+b (1) , (2)
m (2) =W (2) ·(LayerNorm(m (1) ))+b (2) , (3)
P HIN (h,u s ,u e )=W (3) ·concat(m (1) ,m (2) )+b (3) , (4)
wherein equation (1) includes interacting the start interval representation and the end interval representation; the formula (2) comprises interacting the starting interval representation and the ending interval representation, and then integrating the semantic representations of the original training text to obtain a first characteristic; the formula (3) comprises that the first characteristic is subjected to linear processing to obtain a second characteristic; formula (4) includes outputting a start category label or a stop category label after the first feature and the second feature are interacted;
r refers to the feature obtained after the interaction between the starting interval representation and the ending interval, u s Means that the starting interval represents u e The representation of the termination interval, h the semantic representation of the original training text, and m (1) Refers to the first feature, m (2) Refers to said second feature, P HIN (h,u s ,u e ) Refers to the start class label or the end class label, W D 、W (1) 、W (2) 、W (3) Are all weight matrices, b (1) 、b (2) 、b (3) Are all bias parameters.
Optionally, the generating a corresponding start category list and a corresponding end category list based on the start category label and the end category label includes:
P start =argmax eachrow (P HINstart ),
P end =argmax eachrow (P HINend ),
wherein, P start Representing said starting category list, P end Representing said list of termination categories, P HINstart Represents the starting class label, P HINend Indicating the termination category label.
Optionally, the pointer network detection layer includes a starting position determination module and an ending position determination module; the step of inputting the semantic representation of the training text into the pointer network detection layer to obtain the starting interval representation and the ending interval representation of each word in the training text predicted by the pointer network detection layer comprises the following steps:
inputting the semantic representation of the training text into the initial position judging module to obtain the initial interval representation of each word in the training text predicted by the initial position judging module;
and acquiring the real starting interval representation of each word in the training text, inputting the real starting interval representation and the semantic representation of the training text into the termination position judgment module, and acquiring the termination interval representation of each word in the training text predicted by the termination position judgment module.
The invention also provides a social sensing disaster monitoring method based on the segment pointer interaction model, which comprises the following steps:
social media text data are obtained, and preprocessing operation is carried out on the social media text data to obtain a disaster monitoring text;
and inputting the disaster monitoring text into a trained segment pointer interaction model to obtain a disaster event trigger word output by the segment pointer interaction model, wherein the segment pointer interaction model is constructed by the construction method of the segment pointer interaction model.
Optionally, the segment pointer interaction model includes an entity-aware coding layer based on a pre-training language model, a pointer network detection layer, and an interval interaction awareness layer, where the pointer network detection layer includes a start position determination module and an end position determination module; the step of inputting the disaster monitoring text into the trained segment pointer interaction model to obtain the disaster event trigger word output by the segment pointer interaction model comprises the following steps:
inputting the disaster monitoring text into the entity perception coding layer, and obtaining semantic representation of the disaster monitoring text output by the entity perception coding layer;
inputting the semantic representation of the disaster monitoring text into the initial position judgment module to obtain the initial interval representation of each character in the disaster monitoring text predicted by the initial position judgment module;
inputting the starting interval representation of each word in the disaster monitoring text predicted by the starting position determination module and the semantic representation of the disaster monitoring text into the ending position determination module to obtain the ending interval representation of each word in the disaster monitoring text predicted by the ending position determination module;
inputting the starting interval representation and the ending interval representation into the interval interaction perception layer, and performing feature interaction on the starting interval representation and the ending interval representation by the interval interaction perception layer to obtain a starting category label and an ending category label after feature interaction;
and generating a corresponding starting category list and a corresponding ending category list based on the starting category label and the ending category label, and decoding based on the starting category list and the ending category list to obtain a disaster event trigger word.
The invention also provides a social sensing disaster monitoring device based on the segment pointer interaction model, which comprises a computer readable storage medium and a processor, wherein the computer readable storage medium is used for storing a computer program, and the computer program is read by the processor and runs to realize the social sensing disaster monitoring method based on the segment pointer interaction model.
The method and the device realize the detection and identification of the event trigger words in the social media text data by framing the entity interval by utilizing the pointer network detection layer based on the pointer network architecture, can realize the event detection faster than that of the traditional CRF (conditional random field) mode because the pointer network architecture has higher prediction speed, can effectively improve the detection efficiency, has a plurality of irrelevant labels in the social media text data, is spoken, has poor extraction effect on the event trigger words of the data with sparse labels, such as the social media text data, and can also improve the prediction precision by adopting the pointer network architecture. In addition, an interval interactive sensing layer based on a high-speed network is adopted to improve the interactive capacity of the interval information of the trigger words, and after the interval ending position information and the interval starting position information are fully fused, the interval ending position information and the interval starting position information are used for determining a final starting category list and a final ending category list, so that the detection precision of the model is improved, and effective and rapid event detection is realized.
Drawings
FIG. 1 is a diagram of an embodiment of an overall structure of a segment pointer interaction model in a method for constructing the segment pointer interaction model according to the embodiment of the present invention;
FIG. 2 is a schematic diagram of an embodiment of a pointer network detection layer structure of a segment pointer interaction model according to a method for constructing the segment pointer interaction model in the embodiment of the present invention;
FIG. 3 is a schematic diagram of an embodiment of an interaction sensing layer structure between pointer regions of a segment pointer interaction model according to a method for constructing the segment pointer interaction model in the embodiment of the present invention;
FIG. 4 is a flowchart illustrating a method for constructing a segment pointer interaction model according to an embodiment of the present invention;
FIG. 5 is a diagram of an embodiment of a text encoding mode in the method for constructing a segment pointer interaction model according to the embodiment of the present invention;
FIG. 6 is a diagram of an embodiment of an entity-aware coding layer in a method for constructing a segment pointer interaction model according to the embodiment of the present invention;
fig. 7 is a diagram illustrating an embodiment of preprocessing operations in a method for constructing a segment pointer interaction model according to the embodiment of the present invention.
Detailed Description
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in detail below.
The invention provides a construction method of a segment pointer interaction model, and the segment pointer interaction model is used for realizing a social sensing disaster monitoring method. The segment pointer interaction model comprises an entity perception coding layer, a pointer network detection layer and an interval interaction perception layer based on a pre-training language model, wherein the interval interaction perception layer comprises an interaction perception module. FIG. 1 is a diagram illustrating an embodiment of an overall structure of a segment pointer interaction model constructed by the present invention.
In an embodiment of the present invention, as shown in fig. 4, the method for constructing the segment pointer interaction model includes the following steps:
step S1, a training text set is obtained, wherein the training text set is text data obtained after social media text data are subjected to preprocessing operation.
The social media text data refers to text data issued by a user on social media such as a microblog or a WeChat. When a training text set is constructed, preprocessing operation is firstly carried out on social media text data, disaster event labeling is carried out on the preprocessed text data, the preprocessed text data are divided into a training text set and a test text set, the training text set is used for carrying out training operation including the steps S2-S6, and after training is completed, the test text set is used for sending a segment pointer interaction model to evaluate a segment pointer interaction model training result.
The preprocessed text data is subjected to disaster event labeling, for example, if the social media text data is that the palm tree of the road side green belt collapses due to typhoon and the traffic of the related road sections is closed, "collapse" is a trigger word of vegetation loss, "close" is a trigger word of traffic fault, corresponding disaster event labeling is performed for the head and tail characters of the trigger word, the collapse and collapse are respectively labeled as vegetation loss, and the close and close are respectively labeled as traffic fault.
Optionally, the pre-processing operation comprises at least one of: performing deduplication processing on the original social media text data; filtering the original social media text data by adopting a preset keyword template; and filtering the non-event sentences in the original social media text data.
Specifically, because the user autonomy of the social media is strong, the same user may issue a plurality of texts with the same content, and different users may also issue texts with the same content, and in some encoding mechanisms of the social media, such as a microblog encoding mechanism, the user forwarded content and the user issued new content both obtain unique ID identifiers, so that a crawler program distinguished by the user ID may repeatedly crawl a large amount of similar content or the same content, so that the original social media text data needs to be subjected to deduplication processing, and the data processing amount is reduced while disaster detection is not affected. Optionally, in order to efficiently remove repeated text data, the embodiment of the present invention employs a SimHash algorithm, which generates a feature vector of a sentence based on a word included in an original sentence, and does not need external data and a phrase library, thereby greatly saving time and space resources and improving calculation efficiency. Specifically, the SimHash algorithm firstly performs word segmentation on original data, then sets weights for the word segmentation, calculates a Hash value of each word segmentation by a Hash method, multiplies the Hash value by the corresponding weights to obtain a weighted vector of each word, finally adds the weighted vectors of all the words to obtain a weighted vector value of the whole word, performs binarization operation on the weighted vector value of the whole word to obtain a fingerprint vector of the sentence, calculates Hamming distances among different sentence fingerprint vectors to obtain similarity, judges the sentence with the similarity larger than a certain value as a repeated sentence, and performs deduplication processing.
The original social media text data is filtered by using a preset keyword template. Wherein, the keywords in the keyword template comprise disaster situation fact irrelevance texts. Because the social media text data is rich and diverse in content and is not necessarily texts related to disasters, and in the early stage of the meteorological disasters, relevant departments and media issue certain information for early warning according to prediction results, the non-factual texts interfere with data construction of the text. Therefore, a keyword template is set for filtering out texts irrelevant to disaster facts. Optionally, the keyword templates include common sense science popularization type templates and early warning information types, wherein the common sense science popularization type templates include keyword templates such as science popularization knowledge and meteorological knowledge, and the early warning information type templates include keyword templates such as early warning, reminding and prediction.
Filtering non-event sentences in the original social media text data. Because there are quite a few short texts in the social media text data that do not mention the event, the non-event sentences in the social media text data need to be filtered. Here, a Logistic Regression (LR) model may be used to construct a two-classification model for distinguishing event sentences from non-event sentences, and specifically, training data with two-classification event labels may be constructed for model training.
Alternatively, as shown in fig. 7, the preprocessing operation includes: firstly, carrying out duplication elimination processing on the original social media text data, then adopting a preset keyword template to carry out filtering processing on the original social media text data, and finally carrying out filtering processing on non-event sentences in the original social media text data.
Through the preprocessing operation, repeated data, disaster-situation-irrelevant data and non-event data in original social media text data can be screened out, the data processing amount is reduced while the model training effect of the segment pointer interaction model is not influenced, the time consumed by manual labeling is reduced, the interference of irrelevant data is reduced, the data quality is improved, and the training efficiency and the training effect of the segment pointer interaction model are further improved.
Step S2, inputting the training texts in the training text set into the entity perception coding layer, and obtaining the semantic representation of the training texts output by the entity perception coding layer.
After the training text is input into the entity perception coding layer, the semantic representation in the form of the vector is output by the entity perception coding layer. The pre-trained language model may optionally be a BERT model.
Step S3, inputting the semantic representation of the training text into the pointer network detection layer, and obtaining the starting interval representation and the ending interval representation of each word in the training text predicted by the pointer network detection layer.
FIG. 2 is a schematic diagram of an embodiment of a pointer network detection layer structure of a segment pointer interaction model constructed in the present invention. As shown in fig. 2, the pointer network detection layer includes a Start Index Block (SIB) and an End Index Block (EIB). As shown in fig. 2, the Start position determination block includes Dropout processing, Linear linearization processing, and Relu function processing, and the end position determination block includes Start interval representation input Start location, Dropout processing, fusion processing, density 1 processing (first full join processing), Tanh function processing, LayerNorm processing (normalization processing), density 2 processing (second full join processing), and Relu function processing.
Specifically, the start position determination module SIB converts the input sequence flag into c classification for each word, thereby obtaining a start interval representation of each word, where c is the number of event categories. The start interval obtained by the start position decision module SIB is represented as follows:
U start =softmax eachrow (Relu(E·T start +b end ))∈R n*c ,
wherein, U start Is the output probability of the starting position judgment module SIB, E belongs to R n*d Semantic representation output, T, of BERT start ∈R d*c For the weight matrix, d is the BERT vector dimension, n is the number of input words, i.e., sentence length, b end For the bias parameter, eachrow indicates that each word is classified c.
In order to obtain the ending interval representation of each word, the ending position determination module EIB also performs c classification on each word, but needs to input the starting interval representation as additional information to improve entity interval perception, that is, the ending position determination module EIB combines the semantic representation of the training text with the starting interval representation as input. The termination interval obtained by the termination position determination module EIB is expressed as follows:
U end =softmax eachrow (Relu(M·T end +b end ))∈R n*c ,
M=Layernorm(Z),
Z i =tanh(W·concat(U istart ,E i )),
wherein, U end Is the output probability, U, of the EIB istart Indicating the start interval of the ith word, E i Meaning the semantic representation of the ith word, W the weight of Dense1, M the weight of Dense2, T end ∈R d*c Is a weight matrix.
Because the initial position determining module SIB is still in the training phase, the prediction result is still inaccurate, and if an error result is output, the training of the end position determining module EIB is also affected, so that in the training process, in order to avoid the problem of error accumulation propagation, the model convergence is accelerated, and a real initial interval is indicated to be input to the end position determining module EIB. Specifically, the step S3 includes the steps of: inputting the semantic representation of the training text into the initial position judging module to obtain the initial interval representation of each word in the training text predicted by the initial position judging module; and acquiring the real starting interval representation of each word in the training text, inputting the real starting interval representation and the semantic representation of the training text into the termination position judgment module, and acquiring the termination interval representation of each word in the training text predicted by the termination position judgment module.
Step S4, inputting the starting interval representation and the ending interval representation into the interval interaction sensing layer, and performing feature interaction on the starting interval representation and the ending interval representation by the interval interaction sensing layer to obtain a starting category label and an ending category label after the feature interaction.
In order to further improve the detection effect of the entity interval, an interval interaction sensing layer is arranged. In the pointer network detection layer, although the ending position determination module EIB can sense the starting position, when the starting position is predicted, the interval ending information cannot be acquired, which causes a deviation in the prediction stage, and therefore, in the interval interaction sensing layer, feature interaction is performed on the starting interval representation and the ending interval representation, and a starting category label and an ending category label are generated.
Optionally, step S4 includes the steps of: after the initial interval representation and the termination interval representation are interacted, the original semantic representation of the training text is blended to obtain a first feature; performing linear processing on the first characteristic to obtain a second characteristic; and outputting a starting class label or a terminating class label after the first characteristic and the second characteristic are interacted.
Specifically, as shown in fig. 3, the interaction sensing module in the inter-zone interaction sensing layer includes: linear function, LayerNorm. The start category label or the end category label is calculated by:
r=tanh(W D ·concat(u s ,u e )), (1)
m (1) =W (1) ·concat(h,r)+b (1) , (2)
m (2) =W (2) ·(LayerNorm(m (1) ))+b (2) , (3)
P HIN (h,u s ,u e )=W (3) ·concat(m (1) ,m (2) )+b (3) , (4)
wherein equation (1) includes interacting the start interval representation and the end interval representation; the formula (2) comprises the steps of after the starting interval representation and the ending interval representation are interacted, integrating the semantic representation of the original training text to obtain a first characteristic; the formula (3) comprises that the first characteristic is subjected to linear processing to obtain a second characteristic; formula (4) includes outputting a start category label or a stop category label after the first feature and the second feature are interacted;
r refers to the feature obtained after the interaction between the starting interval representation and the ending interval, u s Means that the starting interval represents u e The representation of the termination interval, h the semantic representation of the original training text, and m (1) Refers to the first feature, m (2) Refers to said second feature, P HIN (h,u s ,u e ) Refers to the start class label or the end class label, W D 、W (1) 、W (2) 、W (3) Are all weight matrices, b (1) 、b (2) 、b (3) Are all bias parameters, wherein W D Is a Linear linearized weight matrix.
Wherein, the initial class labels (P) are respectively calculated by adopting the interaction perception modules HINstart ) And a termination category label (P) HINend ) The two are completely symmetrical double-tower structures, only the training targets are different, one is used for judging the head of the interval, the other is used for judging the tail of the interval, and different training data labels can be constructed in the training stage to train corresponding parameters so as to realize the corresponding training targets. For example, if social media text data is that "palm trees in the roadside green belts collapse due to typhoons, traffic of related road sections is closed", "collapse" is a trigger word of vegetation loss, "closed" is a trigger word of traffic faults, corresponding disaster events are labeled for head and tail characters of the trigger word, the "collapse" and the "closure" are respectively labeled as vegetation loss, the "closure" and the "closure" are respectively labeled as traffic faults, and then a training sample of the interactive sensing module for identifying the initial category label is as follows: the text with the triggering word first character label is the text with the double-character disaster situation event label, and other irrelevant characters in the sentence use other labels; the training samples of the interaction perception module for identifying the termination category label are as follows: the text with the triggering word tail character label is the text with the 'collapse' and 'close' two-character disaster situation event label, and other irrelevant characters in the sentence use other labels.
The extraction of the information between the trigger words is carried out through the fragment pointer architecture, the interactive capability of the information between the trigger words is improved by adopting the high way network, and therefore the quick and effective event detection is effectively achieved through the interval interaction.
Step S5, generating a corresponding start category list and an end category list based on the start category label and the end category label, and decoding based on the start category list and the end category list to obtain a disaster event trigger word.
Alternatively, the starting category tag and the terminating category tag may be processed by argmax to obtain a starting category list and a terminating category list, and the argmax processing may be performed to obtain a category tag list with 1-dimensional starting category list and terminating category list. Specifically, the step S5 of generating the corresponding start category list and the corresponding end category list based on the start category label and the end category label includes:
P start =argmax eachrow (P HINstart ),
P end =argmax eachrow (P HINend ),
wherein, P start Represents the starting category list, P end Representing said list of termination categories, P HINstart Represents the starting class label, P HINend Indicating the termination category label.
Decoding the initial category list and the termination category list to obtain a disaster event trigger word, specifically, obtaining the sentence length of an input training text and the maximum length S of the trigger word, for each word in the training text, firstly, obtaining the initial category of the word from the initial category list, judging whether the initial category of the word is a non-event category label, if the initial category of the word is the non-event category label, directly skipping the current word, processing the next word, if the initial category of the word is the event category label, taking the current word as the starting boundary of the current event trigger word, and at the position behind the current word in the termination category list, taking the maximum length S of the trigger word as the limit, searching the termination boundary, if a word with the same category label is found, searching the current event trigger word is ended, and searching the next trigger word is continuously started.
According to the method and the device, the pointer network detection layer based on the pointer network architecture is utilized, the detection and identification of the event trigger words in the social media text data are achieved through the framed entity interval, the pointer network architecture has a high prediction speed, the event detection which is faster than that of a traditional CRF mode can be achieved, the detection efficiency can be effectively improved, the social media text data have a large number of irrelevant labels and are spoken, the traditional CRF mode has a poor effect of extracting the event trigger words from the social media text data with sparse labels, and the prediction precision can be improved by adopting the pointer network architecture. In addition, the interactive capacity of the interval information of the trigger words is improved by adopting an interval interactive perception layer based on the high way network, and after the interval ending position information and the interval starting position information are fully fused, the interval ending position information and the interval starting position information are used for determining a final starting category list and a final ending category list, so that the detection precision of the model is improved, and effective and rapid event detection is realized.
Optionally, the pre-processing operation comprises the steps of:
extracting entity information from the social media text based on a preset word segmentation algorithm, wherein the entity information comprises an entity type of an extracted entity word and position information of the entity word in the social media text; and attaching the entity type and the position information of the entity words extracted from the social media text to be used as the training text.
Due to the complexity of the language, different events may have the same event trigger for indicating different event types, for example, S1 "typhoon attacks the XX area of XX city positively, causing green trees of a plurality of main roads in the XX area to collapse and block traffic", S2 "scene of incident, main building of XX building is blown down by strong wind, main building collapses", wherein the collapse in S1 indicates "traffic problem", and the collapse in S2 indicates "civil problem". It is therefore necessary to focus on the ambiguity problem of the trigger word.
Through observation of disaster field data, objective entities of nouns, place names and person names in a disaster text can play an auxiliary judgment role in the category of the event, so that the ambiguity problem of trigger words is solved. Therefore, a preset word segmentation algorithm such as Jieba word segmentation is used for performing word segmentation on the social media text in one step, entity information in the social media text is extracted, a corresponding entity type (the entity type is a place name, an object name and the like identified by Jieba) and a subscript (the subscript is position information of the entity) corresponding to the text are recorded, and the entity type and the position information of the extracted entity word are attached to the social media text to serve as a training text.
Further, the pre-training language model is a BERT model; the step S2 includes an input encoding step when the entity-aware encoding layer is input, and specifically includes:
and coding the social media texts into an upper sentence and a lower sentence, wherein one sentence corresponds to the sequence coding of the full social media texts, and the other sentence corresponds to the coding of the entity type and the position information of the entity words extracted from the social media texts.
In order to integrate entity information into a pre-training language model in a pluggable manner, the invention is different from the traditional text information input mode such as a text coding mode shown in fig. 5(a), wherein a text coding mode shown in fig. 5(b) is adopted, in the fig. 5(a), a text is coded in a [ cls ] text [ sep ] form, and the position of each word is coded in sequence, in the fig. 5(b), a text is coded in a [ cls ] text [ sep ] text-entry [ sep ] form, wherein the text-entry is used for coding an entity type, and the position coding adopts the position of a corresponding entity word.
Further, as shown in fig. 6, the input code comprises three parts, and word codes are used for representing semantics; segment codes are used for learning upper and lower sentence information; the position coding is used for providing position information, the invention changes the original input mode of the BERT model, and changes the single sentence input in the [ cls ] text [ sep ] form into the upper and lower sentence form. In one embodiment, as shown in fig. 6, the upper sentence corresponds to the sequential coding of the full text of the social media text, the lower sentence corresponds to the coding of the entity type and the position information of the entity word extracted from the social media text, the coding is in the form of [ cls ] text [ sep ] text-entry [ sep ], the position coding is modified, the upper sentence is sequentially represented in position, and the position corresponding to each entity in the lower sentence is modified. In another embodiment, the upper sentence corresponds to the encoding of the entity type and location information of the entity words extracted from the social media text, and the lower sentence corresponds to the sequential encoding of the full text of the social media text.
By changing the traditional text coding mode, the event related entity word information is extracted and merged into the text coding, the entity information in the sentence is fully utilized, the interference of trigger word ambiguity in the event detection problem can be reduced, and the prediction precision is effectively improved.
Optionally, before the step of extracting entity information from the social media text based on the preset word segmentation algorithm, the preprocessing operation further includes the following steps:
obtaining original social media text data, and processing the original social media text data by adopting at least one of the following operations: performing deduplication processing on the original social media text data; filtering the original social media text data by adopting a preset keyword template, wherein keywords in the keyword template comprise disaster situation fact irrelevance texts; and filtering the non-event sentences in the original social media text data. The details of the three operations are described in detail above, and are not repeated herein.
Optionally, the invention trains the segment pointer interaction model using a cross entropy loss function. The penalty function of the segment pointer interaction model is expressed as follows:
L start =CE(P start ,Y start ),
L end =CE(P end ,Y end ),
L=L start +L end ,
wherein L is start Loss function, L, of the decision block for the starting position end A finger termination position judgment module, L indicates the total loss function of the segment pointer interaction model, P start Indicates the section start position, Y, predicted by the start position decision module start Refers to the actual starting position of the interval, P end Section end position, Y, predicted by the finger end position determination module end Refers to the actual interval end position.
The invention further provides a social sensing disaster monitoring method based on the segment pointer interaction model. In an embodiment of the invention, the social sensing disaster monitoring method based on the segment pointer interaction model comprises the following steps:
and acquiring social media text data, and performing preprocessing operation on the social media text data to obtain a disaster monitoring text.
And inputting the disaster monitoring text into a trained segment pointer interaction model to obtain a disaster event trigger word output by the segment pointer interaction model, wherein the segment pointer interaction model is constructed by the construction method of the segment pointer interaction model.
The social media text data refers to text data issued by a user on social media such as microblogs or WeChats, real-time social media text data can be continuously crawled through a crawler, and a trained segment pointer interaction model is input to perform event detection. The pre-processing operation includes at least one of: performing deduplication processing on the original social media text data; and filtering the original social media text data by adopting a preset keyword template. The details of the above operation are described in detail above, and are not repeated herein.
Optionally, during actual use after the segment pointer interaction model is trained, the latest historical data is input into the trained segment pointer interaction model for incremental training to enhance the performance of the model, wherein the latest historical data may refer to the historical data of the previous day or the previous days.
Optionally, the segment pointer interaction model includes an entity-aware coding layer based on a pre-training language model, a pointer network detection layer, and an interval interaction-aware layer, where the pointer network detection layer includes a start position determination module and an end position determination module. The step of inputting the disaster monitoring text into the trained segment pointer interaction model to obtain the disaster event trigger word output by the segment pointer interaction model comprises the following steps:
inputting the disaster monitoring text into the entity perception coding layer, and obtaining semantic representation of the disaster monitoring text output by the entity perception coding layer;
inputting the semantic representation of the disaster monitoring text into the initial position judging module to obtain the initial interval representation of each word in the disaster monitoring text predicted by the initial position judging module;
inputting the starting interval representation of each word in the disaster monitoring text predicted by the starting position judgment module and the semantic representation of the disaster monitoring text into the ending position judgment module to obtain the ending interval representation of each word in the disaster monitoring text predicted by the ending position judgment module;
inputting the starting interval representation and the ending interval representation into the interval interaction sensing layer, and performing feature interaction on the starting interval representation and the ending interval representation by the interval interaction sensing layer to obtain a starting category label and an ending category label after feature interaction;
and generating a corresponding starting category list and a corresponding ending category list based on the starting category label and the ending category label, and decoding based on the starting category list and the ending category list to obtain a disaster event trigger word.
The relevant explanation of the above steps has been detailed in the foregoing, and different from the above-mentioned segment pointer interaction model construction process/training process, in the actual use process of the segment pointer interaction model trained here, the input of the termination position determination module is the semantic representation of the start interval representation and the disaster monitoring text predicted by the start position determination module, rather than the actual semantic representation of the start interval representation and the disaster monitoring text.
The disaster monitoring text is input to the trained segment pointer interaction model to obtain the disaster event trigger words output by the segment pointer interaction model, the segment pointer interaction model utilizes a pointer network detection layer based on a pointer network architecture to realize the detection and identification of the event trigger words in the social media text data through framing an entity interval, and the pointer network architecture has higher prediction speed, so that the event detection is faster than that of the traditional CRF mode, the detection efficiency can be effectively improved, the social media text data has many irrelevant labels and is spoken, the traditional CRF mode has poor extraction effect on the event trigger words of the data with sparse labels, namely the social media text data, and the prediction accuracy can be improved by adopting the pointer network architecture. In addition, the interactive capacity of the interval information of the trigger words is improved by adopting an interval interactive perception layer based on the high way network, and after the interval ending position information and the interval starting position information are fully fused, the interval ending position information and the interval starting position information are used for determining a final starting category list and a final ending category list, so that the detection precision of the model is improved, and effective and rapid event detection is realized.
In an embodiment of the present invention, the social sensing disaster monitoring apparatus based on the segment pointer interaction model includes a computer readable storage medium storing a computer program and a processor, and when the computer program is read and executed by the processor, the social sensing disaster monitoring method based on the segment pointer interaction model is implemented. Compared with the prior art, the social sensing disaster monitoring device based on the segment pointer interaction model has the advantages that the social sensing disaster monitoring device based on the segment pointer interaction model has the same advantages as the social sensing disaster monitoring method based on the segment pointer interaction model, and the details are omitted.
The reader should understand that in the description of this specification, reference to the description of the terms "one embodiment," "some embodiments," "an example," "a specific example" or "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Moreover, various embodiments or examples and features of various embodiments or examples described in this specification can be combined and combined by one skilled in the art without being mutually inconsistent.
Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention, and that variations, modifications, substitutions and alterations can be made to the above embodiments by those of ordinary skill in the art within the scope of the present invention.
Claims (10)
1. A construction method of a segment pointer interaction model is characterized in that the segment pointer interaction model comprises an entity perception coding layer, a pointer network detection layer and an interval interaction perception layer based on a pre-training language model; the construction method of the segment pointer interaction model comprises the following steps:
acquiring a training text set, wherein the training text set is text data obtained by preprocessing social media text data;
inputting the training texts in the training text set into the entity perception coding layer to obtain semantic representation of the training texts output by the entity perception coding layer;
inputting the semantic representation of the training text into the pointer network detection layer to obtain the starting interval representation and the ending interval representation of each word in the training text predicted by the pointer network detection layer;
inputting the starting interval representation and the ending interval representation into the interval interaction perception layer, and performing feature interaction on the starting interval representation and the ending interval representation by the interval interaction perception layer to obtain a starting category label and an ending category label after feature interaction;
and generating a corresponding starting category list and a corresponding terminating category list based on the starting category label and the terminating category label, and decoding based on the starting category list and the terminating category list to obtain the disaster event trigger word.
2. The method for building a segment pointer interaction model according to claim 1, wherein the preprocessing operation comprises the steps of:
extracting entity information from the social media text based on a preset word segmentation algorithm, wherein the entity information comprises an entity type of an extracted entity word and position information of the entity word in the social media text;
attaching the entity type and the position information of the entity word extracted from the social media text to be used as the training text;
the pre-training language model is a BERT model; the step of inputting the training texts in the training text set into the entity-aware coding layer to obtain semantic representations of the training texts output by the entity-aware coding layer includes an input coding step when the semantic representations are input into the entity-aware coding layer, which specifically includes:
and coding the social media texts into an upper sentence and a lower sentence, wherein one sentence corresponds to the sequence coding of the full social media texts, and the other sentence corresponds to the coding of the entity type and the position information of the entity words extracted from the social media texts.
3. The method for constructing a segment pointer interaction model according to claim 2, wherein, before the step of extracting entity information from the social media text based on the preset word segmentation algorithm, the preprocessing operation further comprises the following steps:
obtaining original social media text data, and processing the original social media text data by adopting at least one of the following operations:
performing deduplication processing on the original social media text data;
filtering the original social media text data by adopting a preset keyword template, wherein keywords in the keyword template comprise disaster fact irrelevant texts;
and filtering the non-event sentences in the original social media text data.
4. The method for constructing a segment pointer interaction model according to claim 1, wherein the step of inputting the start interval representation and the end interval representation into the interval interaction sensing layer, and performing feature interaction on the start interval representation and the end interval representation by the interval interaction sensing layer to obtain a start category label and an end category label after feature interaction comprises the steps of:
after the initial interval representation and the termination interval representation are interacted, the original semantic representation of the training text is blended to obtain a first feature;
performing linear processing on the first characteristic to obtain a second characteristic;
and outputting a starting class label or a terminating class label after the first characteristic and the second characteristic are interacted.
5. The method for constructing a segment pointer interaction model according to claim 4, wherein the inputting the start interval representation and the end interval representation into the interval interaction sensing layer, and performing feature interaction on the start interval representation and the end interval representation by the interval interaction sensing layer to obtain a start category label and an end category label after feature interaction comprises:
r=tanh(W D ·concat(u s ,u e )), (1)
m (1) =W (1) ·concat(h,r)+b (1) , (2)
m (2) =W (2) ·(LayerNorm(m (1) ))+b (2) , (3)
P HIN (h,u s ,u e )=W (3) ·concat(m (1) ,m (2) )+b (3) , (4)
wherein equation (1) includes interacting the start interval representation and the end interval representation; the formula (2) comprises interacting the starting interval representation and the ending interval representation, and then integrating the semantic representations of the original training text to obtain a first characteristic; the formula (3) comprises that the first characteristic is subjected to linear processing to obtain a second characteristic; formula (4) includes outputting a start category label or a stop category label after the first feature and the second feature are interacted;
r refers to the feature obtained after the interaction between the starting interval representation and the ending interval, u s Means that the starting interval represents u e Means the termination ofInterval representation, h denotes the original semantic representation of the training text, m (1) Refers to the first feature, m (2) Refers to said second feature, P HIN (h,u s ,u e ) Refers to the start class label or the end class label, W D 、W (1) 、W (2) 、W (3) Are all weight matrices, b (1) 、b (2) 、b (3) Are all bias parameters.
6. The method of constructing a segment pointer interaction model of claim 4 or 5, wherein said generating a corresponding start category list and a corresponding end category list based on said start category label and said end category label comprises:
P start =argmax eachrow (P HINstart ),
P end =argmax eachrow (P HINend ),
wherein, P start Representing said starting category list, P end Representing said list of termination categories, P HINstart Represents the starting class label, P HINend Representing the termination category label.
7. The method for constructing a segment pointer interaction model according to claim 1, wherein the pointer network detection layer comprises a starting position decision module and an ending position decision module; the step of inputting the semantic representation of the training text into the pointer network detection layer to obtain the starting interval representation and the ending interval representation of each word in the training text predicted by the pointer network detection layer comprises the following steps:
inputting the semantic representation of the training text into the initial position judging module to obtain the initial interval representation of each word in the training text predicted by the initial position judging module;
and acquiring the real starting interval representation of each word in the training text, inputting the real starting interval representation and the semantic representation of the training text into the ending position judgment module, and acquiring the ending interval representation of each word in the training text predicted by the ending position judgment module.
8. A social sensing disaster monitoring method based on a segment pointer interaction model is characterized by comprising the following steps:
social media text data are obtained, and preprocessing operation is carried out on the social media text data to obtain a disaster monitoring text;
inputting the disaster monitoring text into a trained segment pointer interaction model, and obtaining a disaster event trigger word output by the segment pointer interaction model, wherein the segment pointer interaction model is constructed by the construction method of the segment pointer interaction model according to any one of claims 1 to 7.
9. The social sensing disaster monitoring method based on the segment pointer interaction model as claimed in claim 8, wherein the segment pointer interaction model includes an entity sensing coding layer based on a pre-training language model, a pointer network detection layer and an interval interaction sensing layer, the pointer network detection layer includes a start position determination module and an end position determination module; the step of inputting the disaster monitoring text into the trained segment pointer interaction model to obtain the disaster event trigger word output by the segment pointer interaction model comprises the following steps:
inputting the disaster monitoring text into the entity perception coding layer, and obtaining semantic representation of the disaster monitoring text output by the entity perception coding layer;
inputting the semantic representation of the disaster monitoring text into the initial position judgment module to obtain the initial interval representation of each character in the disaster monitoring text predicted by the initial position judgment module;
inputting the starting interval representation of each word in the disaster monitoring text predicted by the starting position determination module and the semantic representation of the disaster monitoring text into the ending position determination module to obtain the ending interval representation of each word in the disaster monitoring text predicted by the ending position determination module;
inputting the starting interval representation and the ending interval representation into the interval interaction sensing layer, and performing feature interaction on the starting interval representation and the ending interval representation by the interval interaction sensing layer to obtain a starting category label and an ending category label after feature interaction;
and generating a corresponding starting category list and a corresponding terminating category list based on the starting category label and the terminating category label, and decoding based on the starting category list and the terminating category list to obtain the disaster event trigger word.
10. A social sensing disaster monitoring device based on a segment pointer interaction model, which comprises a computer readable storage medium storing a computer program and a processor, wherein the computer program is read by the processor and when the computer program is executed by the processor, the social sensing disaster monitoring device based on the segment pointer interaction model according to claim 8 or 9 is implemented.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210374738.9A CN114943221A (en) | 2022-04-11 | 2022-04-11 | Construction method of segment pointer interaction model and social sensing disaster monitoring method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210374738.9A CN114943221A (en) | 2022-04-11 | 2022-04-11 | Construction method of segment pointer interaction model and social sensing disaster monitoring method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114943221A true CN114943221A (en) | 2022-08-26 |
Family
ID=82906664
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210374738.9A Pending CN114943221A (en) | 2022-04-11 | 2022-04-11 | Construction method of segment pointer interaction model and social sensing disaster monitoring method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114943221A (en) |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113076468A (en) * | 2021-04-27 | 2021-07-06 | 华东理工大学 | Nested event extraction method based on domain pre-training |
CN113283244A (en) * | 2021-07-20 | 2021-08-20 | 湖南达德曼宁信息技术有限公司 | Pre-training model-based bidding data named entity identification method |
CN113312916A (en) * | 2021-05-28 | 2021-08-27 | 北京航空航天大学 | Financial text event extraction method and device based on triggered word morphological learning |
CN113779988A (en) * | 2021-09-07 | 2021-12-10 | 科大国创云网科技有限公司 | Method for extracting process knowledge events in communication field |
US20210406476A1 (en) * | 2020-06-30 | 2021-12-30 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Method, electronic device, and storage medium for extracting event from text |
CN113901209A (en) * | 2021-09-15 | 2022-01-07 | 昆明理工大学 | Chinese cross-language event detection method based on type perception |
CN114239566A (en) * | 2021-12-14 | 2022-03-25 | 公安部第三研究所 | Method, device and processor for realizing two-step Chinese event accurate detection based on information enhancement and computer readable storage medium thereof |
CN114254028A (en) * | 2021-12-20 | 2022-03-29 | 北京百度网讯科技有限公司 | Event attribute extraction method and device, electronic equipment and storage medium |
-
2022
- 2022-04-11 CN CN202210374738.9A patent/CN114943221A/en active Pending
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20210406476A1 (en) * | 2020-06-30 | 2021-12-30 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Method, electronic device, and storage medium for extracting event from text |
CN113076468A (en) * | 2021-04-27 | 2021-07-06 | 华东理工大学 | Nested event extraction method based on domain pre-training |
CN113312916A (en) * | 2021-05-28 | 2021-08-27 | 北京航空航天大学 | Financial text event extraction method and device based on triggered word morphological learning |
CN113283244A (en) * | 2021-07-20 | 2021-08-20 | 湖南达德曼宁信息技术有限公司 | Pre-training model-based bidding data named entity identification method |
CN113779988A (en) * | 2021-09-07 | 2021-12-10 | 科大国创云网科技有限公司 | Method for extracting process knowledge events in communication field |
CN113901209A (en) * | 2021-09-15 | 2022-01-07 | 昆明理工大学 | Chinese cross-language event detection method based on type perception |
CN114239566A (en) * | 2021-12-14 | 2022-03-25 | 公安部第三研究所 | Method, device and processor for realizing two-step Chinese event accurate detection based on information enhancement and computer readable storage medium thereof |
CN114254028A (en) * | 2021-12-20 | 2022-03-29 | 北京百度网讯科技有限公司 | Event attribute extraction method and device, electronic equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110516067B (en) | Public opinion monitoring method, system and storage medium based on topic detection | |
CN108897857B (en) | Chinese text subject sentence generating method facing field | |
CN103544255B (en) | Text semantic relativity based network public opinion information analysis method | |
CN110334178B (en) | Data retrieval method, device, equipment and readable storage medium | |
CN111783394B (en) | Training method of event extraction model, event extraction method, system and equipment | |
CN112256939B (en) | Text entity relation extraction method for chemical field | |
CN110347894A (en) | Knowledge mapping processing method, device, computer equipment and storage medium based on crawler | |
WO2021051518A1 (en) | Text data classification method and apparatus based on neural network model, and storage medium | |
CN113806563B (en) | Architect knowledge graph construction method for multi-source heterogeneous building humanistic historical material | |
CN112650848A (en) | Urban railway public opinion information analysis method based on text semantic related passenger evaluation | |
CN109657058A (en) | A kind of abstracting method of notice information | |
CN114579739B (en) | Topic detection and tracking method for text data stream | |
CN113449111B (en) | Social governance hot topic automatic identification method based on time-space semantic knowledge migration | |
CN113065356B (en) | IT equipment operation and maintenance fault suggestion processing method based on semantic analysis algorithm | |
CN113076483A (en) | Case element heteromorphic graph-based public opinion news extraction type summarization method | |
CN115017425B (en) | Location search method, location search device, electronic device, and storage medium | |
CN112270187A (en) | Bert-LSTM-based rumor detection model | |
CN117574898A (en) | Domain knowledge graph updating method and system based on power grid equipment | |
CN112883182A (en) | Question-answer matching method and device based on machine reading | |
CN116737922A (en) | Tourist online comment fine granularity emotion analysis method and system | |
CN117235243A (en) | Training optimization method for large language model of civil airport and comprehensive service platform | |
CN114764463A (en) | Internet public opinion event automatic early warning system based on event propagation characteristics | |
CN116955818A (en) | Recommendation system based on deep learning | |
CN114638222B (en) | Natural disaster data classification method and model training method and device thereof | |
CN113177164B (en) | Multi-platform collaborative new media content monitoring and management system based on big data |
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 |