CN109614612A - A kind of Chinese text error correction method based on seq2seq+attention - Google Patents
A kind of Chinese text error correction method based on seq2seq+attention Download PDFInfo
- Publication number
- CN109614612A CN109614612A CN201811441697.0A CN201811441697A CN109614612A CN 109614612 A CN109614612 A CN 109614612A CN 201811441697 A CN201811441697 A CN 201811441697A CN 109614612 A CN109614612 A CN 109614612A
- Authority
- CN
- China
- Prior art keywords
- character
- layer
- vector
- moment
- lstm
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/232—Orthographic correction, e.g. spell checking or vowelisation
-
- 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
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield 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
- G06N3/04—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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Data Mining & Analysis (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Machine Translation (AREA)
Abstract
The present invention relates to a kind of Chinese text error correction method based on seq2seq+attention, belong to the research category of the quality of data, it is related to RNN, two-way RNN, LSTM, seq2seq, the technical fields such as attention mechanism construct seq2seq+attention neural network model mainly for communication equipment record of examination, model training is carried out using Adam optimization method, carries out error checking tasks using trained model.Used neural network model can be widely applied to the text error correction of other field in this method, avoid the redesign of model to a certain extent.
Description
Technical field
The invention belongs to the technical field of Chinese text error correction, it is related specifically to generate in Power Communication Management System logical
Believe the field of error correction of overhaul of the equipments record.
Background technique
The main study subject that the field is related to, key technology and practical application value specifically include that
Power Communication Management System: it is the electric power dedicated communications network system as smart grid important support, is general headquarters
With provincial company " two-stage deployment ", general headquarters, branch, provincial company, company, cities and counties " level Four application " communications management system " SG-
TMS".It is carried forward vigorously by the project construction of standardization and normalization and to system functionization, " SG-TMS " depth is melted
In the routine work for entering tens of thousands of power communication professionals, and acquire comprehensively the construction in the past few years of tens of thousands of equipment, operation,
Data are managed, accumulative magnanimity Electric Power Communication Data and numerous external system data, common data have been formed together development
The basis of big data analysis.
Overhaul of the equipments record: a large amount of overhaul datas, side have been had accumulated in the Information Management System of smart grid communication
Formula data, log data also have many classes wherein the structural data of existing specification such as overhauls type, execution date
Like the semi-structured data of log one kind, there are also many similar routing modes descriptions, " three arrange a case " document, picture etc. are non-
Structural data.By the in-depth analysis and excavation to these processes and findings data, management rule can be summed up, to existing
System and way to manage optimizes and reasonable arrangement.It can also be realized by big data means to the method for operation, " three arrange one
The automatic first trial of machine of the procedures such as case " work, intelligent functions of filling a vacancy etc. to the automatic auxiliary error correction of logging, reduce pipe
Personnel labor intensity is managed, hoisting way, maintenance examination & approval efficiency and record are normative.
Based on word granularity: mainly there are two reasons: first, due to including wrong word in overhaul of the equipments record or existing scarce
Situation is lost, causes word segmentation result inaccurate, so error checking tasks are not suitable for carrying out in word granularity;Second, in given fixation
In the case where vocabulary, the error checking tasks based on word can not handle the word of OOV (out ofvocabulary).
RNN:RNN is a kind of sequence link model, captures dynamic sequence by the circulation in network node.Before standard
It is different to present neural network, RNN can retain the status information of the contextual window of random length.Although RNN is difficult to train, and
Millions of a parameters, but the network architecture are generally comprised, the latest developments in terms of optimisation technique and parallel computation allow them to
Successfully learnt on a large scale.
Two-way RNN: in classical Recognition with Recurrent Neural Network, state is one-way transmission from front to back.However, asking at some
In topic, the state of the output at current time not only and before has relationship, also related to state later.At this moment it just needs two-way
RNN is come such issues that solve.For example, predicting the word lacked in a sentence not only in text error checking tasks of the invention
It needs to judge according to above, it is also desirable to which, in conjunction with subsequent content, at this moment two-way RNN can play its effect.
LSTM:Long Short-Term Memory is shot and long term memory network.RNN is in long-term dependence (the time sequence of processing
The farther away node of distance on column) when can get into enormous difficulties, can be related to when because calculating connection between farther away node
And the multiple multiplication of Jacobian matrix, this can bring gradient disappear or gradient explode the problem of.In order to solve this problem, Sepp
Hochreiter et al.[3]LSTM model is proposed, by increasing input threshold, thresholding and output thresholding are forgotten, so that self-loopa
Weight be variation, so in the case where model parameter is fixed, the integral scale of different moments can dynamically change,
The problem of being expanded so as to avoid gradient disappearance or gradient.
Seq2seq:seq2seq is the network of an Encoder-Decoder structure, its input is a sequence, defeated
Out and a sequence, the vector that the signal sequence of a variable-length becomes regular length is expressed in Encoder,
The vector of this regular length is become the signal sequence of the target of variable-length by Decoder.
Attention mechanism: using the neural network model of coder-decoder structure need by list entries must
It wants information to be expressed as the vector of a regular length, and is then difficult to retain whole necessary informations when list entries is very long, especially
It is when the length of list entries than training data concentrate it is longer when.DzmitryBahdanau et al. uses attention machine
It makes one word Shi Douhui of every generation and finds out a maximally related set of words therewith in list entries, model is according to current later
Context vector and it is all before the word that generates predict next target word
Summary of the invention
Error detection and error correction for communication equipment record of examination, the invention proposes based on attention mechanism
Seq2seq neural network model.The step of realizing record of examination error correction is as follows:
A kind of Chinese text error correction method based on seq2seq+attention, which comprises the following steps:
Step 1, Text Pretreatment: the record of examination being primarily based in python reading database extracts in document files
All the elements, then carry out Chinese subordinate sentence using regular expression and operate, result is stored in text file, every a line is corresponding
One sentence, at the same by the correct text manually marked be stored in another text file and original document correspond;Record
Proprietary symbol and common Chinese character table in lower field of power communication collectively constitute character list;
Step 2, the seq2seq neural network model based on attention is constructed, is specifically included:
Step 2.1, Encoder module layer, including Layer and M layers of two-way LSTM of Embedding are constructed, in which:
The input of layer one, Embedding Layer is that the one-hot of current character is encoded, and one-hot coding can basis
The character list formed in step 1 obtains;The output of Embedding Layer be current character word vector i.e.:
et=ET·xt
Wherein xtIt is the one-hot coding of t moment input character, is v dimensional vector, v is in the character list that step 1 obtains
Character sum;E is character vector matrix, is v × d dimension matrix, during specific implementation, d takes the number between 100-200
Word, d represent the dimension of each character vector, and matrix E is the parameter of model, are obtained by training;etIt is t moment input character
Word vector;
During specific implementation, using tf.nn.embedding_lookup function in TensorFlow, character is obtained
Vector;
Basic unit in layer two, M layers of two-way LSTM is LSTM, and the hidden state calculation formula of jth layer t moment LSTM is such as
Under:
Wherein, It is initialized as null vector, function σ (x)=1/ (1+e-x), function tanh (x)=(ex-e-x)/(ex+e-x),It is model parameter, is obtained by training, ⊙
Indicate that corresponding element is multiplied;
Calculation formula based on LSTM obtains: the hidden state vector propagated from front to backThat propagates from back to front is hidden
Hide state vectorThen in jth layer t moment, the output of Bi-LSTM is
During specific implementation, the function BasicLSTMCell in TensorFlow can be used to realize;
Step 2.2, Decoder module layer is constructed, Decoder module layer is one M layers unidirectional LSTM language model structure,
Each layer of initial state vector is all from Encoder;The output for representing M layers of Decoder, converts by softmax, obtains
To the probability of each character, calculation formula is as follows:
Calculate loss function:
Wherein, ytIt is one-hot coding, represents the character of t moment output, P is output sequence length;
Step 2.3, it constructs attention module layer: obtaining the hidden of the last layer all moment in Encoder module layer
Hide state vectorQ is the sum for inputting character;The first layer t-1 moment is obtained in Decoder module
Hidden state vectorThe calculation formula of attention vector is as follows:
α=softmax (w)
In calculating Decoder module layer when each layer of t moment of hidden state vector, β is added;
Step 3, carry out model training using Adam optimization method: after neural network model constructs successfully, data are from defeated
Enter the whole flow process to output, exist in the form of TensorFlow falls into a trap nomogram, directly utilizes
Tf.train.AdamOptimizer () .minimize () is iterated training, finds parameter optimal value;
Step 4, error checking tasks are carried out, specifically:
In the Decoder stage, deduction process is carried out using beam search, beam size is 2, selects two every time generally
The highest character of rate when meeting full stop<EOS>, stops inferring, obtains output sequence as the input predicted next time.
Therefore, the present invention has the advantage that
(1) with existing rule-based error correction scheme, depth model can extract feature to avoid artificial end to end, subtract
Few labor workload.Moreover, RNN series model is strong to text task capability of fitting.
(2) attention mechanism is added in seq2seq model, more preferable for long text effect, model is easier to restrain,
Effect is more preferable.
Detailed description of the invention
Fig. 1 is neural network overall framework.
Specific embodiment
When it is implemented, technical solution provided by the present invention can be real using computer software technology by those skilled in the art
Existing automatic running process.Below in conjunction with drawings and examples the present invention will be described in detail technical solution.
Step 1: Text Pretreatment
Using the record of examination in the related tool reading database in python, extract all interior in document files
Hold, then carries out Chinese subordinate sentence using regular expression and operate, result is stored in text file, the corresponding sentence of every a line
Son, at the same by the correct text manually marked be stored in another text file and original document correspond.Record electric power
Proprietary symbol and common Chinese character table in the communications field collectively constitute character list used in the present invention.
Step 2: seq2seq neural network model of the building based on attention
(1) Encoder module
Encoder module mainly includes two parts: Layer and M layers of two-way LSTM of Embedding.
The input of EmbeddingLayer is that the one-hot of current character is encoded, and one-hot coding can be according to step 1
The character list of middle formation obtains.The output of Embedding Layer be current character word vector i.e.:
et=ET·xt
Wherein xtIt is the one-hot coding of t moment input character, is v dimensional vector, v is in the character list that step 1 obtains
Character sum.E is character vector matrix, is v × d dimension matrix, during specific implementation, d takes the number between 100-200
Word, d represent the dimension of each character vector, and matrix E is the parameter of model, are obtained by training.etIt is t moment input character
Word vector.
During specific implementation, using tf.nn.embedding_lookup function in TensorFlow, character is obtained
Vector.
Basic unit in M layers of two-way LSTM is LSTM, and the hidden state calculation formula of jth layer t moment LSTM is as follows:
Wherein, It is initialized as null vector, function σ (x)=1/ (1+e-x), function tanh (x)=(ex-e-x)/(ex+e-x),It is model parameter, is obtained by training, ⊙
Indicate that corresponding element is multiplied.
Calculation formula based on above-mentioned LSTM obtains: the hidden state vector propagated from front to backIt propagates from back to front
Hidden state vectorThen in jth layer t moment, the output of Bi-LSTM is
During specific implementation, the function BasicLSTMCell in TensorFlow can be used to realize.
(2) Decoder module
Decoder module is one M layers unidirectional LSTM language model structure, and each layer of initial state vector is all from
Encoder。The output for representing M layers of Decoder, converts by softmax, obtains the probability of each character, calculation formula
It is as follows:
Calculate loss function:
Wherein, ytIt is one-hot coding, represents the character of t moment output, P is output sequence length.
(3) attention module
The hidden state vector at the last layer all moment is obtained in Encoder moduleQ is input
The sum of character.The hidden state vector at first layer t-1 moment is obtained in Decoder moduleAttention vector
Calculation formula is as follows:
α=softmax (w)
In calculating Decoder module when each layer of t moment of hidden state vector, β is added.
Step 3: model training is carried out using Adam optimization method
After neural network model constructs successfully, data are fallen into a trap from the whole flow process for being input to output with TensorFlow
The form of nomogram exists, and is directly iterated training using tf.train.AdamOptimizer () .minimize (), finds
Parameter optimal value.
Step 4: carrying out error checking tasks
In the Decoder stage, deduction process is carried out using beam search, beam size is 2, selects two every time generally
The highest character of rate when meeting full stop<EOS>, stops inferring, obtains output sequence as the input predicted next time.
Specific embodiment described herein is only an example for the spirit of the invention.The neck of technology belonging to the present invention
The technical staff in domain can make an amendment or supplement or be substituted in a similar manner to described specific embodiment, but can't be inclined
From spirit or beyond the scope defined by the appended claims of the invention.
Claims (1)
1. a kind of Chinese text error correction method based on seq2seq+attention, which comprises the following steps:
Step 1, Text Pretreatment: the record of examination being primarily based in python reading database extracts the institute in document files
There is content, then carries out Chinese subordinate sentence using regular expression and operate, result is stored in text file, every a line is one corresponding
Sentence, at the same by the correct text manually marked be stored in another text file and original document correspond;Record electricity
Proprietary symbol and common Chinese character table in the power communications field collectively constitute character list;
Step 2, the seq2seq neural network model based on attention is constructed, is specifically included:
Step 2.1, Encoder module layer, including Layer and M layers of two-way LSTM of Embedding are constructed, in which:
The input of layer one, Embedding Layer is that the one-hot of current character is encoded, and one-hot coding can be according to step
The character list formed in 1 obtains;EmbeddingThe output of Layer be current character word vector i.e.:
et=ET·xt
Wherein xtIt is the one-hot coding of t moment input character, is v dimensional vector, v is the word in the character list that step 1 obtains
Symbol sum;E is character vector matrix, is v × d dimension matrix, during specific implementation, d takes the number between 100-200, d generation
The dimension of each character vector of table, matrix E are the parameters of model, are obtained by training;etBe t moment input character word to
Amount;
During specific implementation, using tf.nn.embedding_lookup function in TensorFlow, character vector is obtained;
Basic unit in layer two, M layers of two-way LSTM is LSTM, and the hidden state calculation formula of jth layer t moment LSTM is as follows:
Wherein, It is initialized as null vector, function σ (x)=1/ (1+e-x), function tanh (x)=(ex-e-x)/(ex
+e-x),It is model parameter, is obtained by training, ⊙ expression pair
Answer element multiplication;
Calculation formula based on LSTM obtains: the hidden state vector propagated from front to backThe hidden state propagated from back to front
VectorThen in jth layer t moment, the output of Bi-LSTM is
During specific implementation, the function BasicLSTMCell in TensorFlow can be used to realize;
Step 2.2, Decoder module layer is constructed, Decoder module layer is one M layers unidirectional LSTM language model structure, each
Layer initial state vector is all from Encoder;The output for representing M layers of Decoder, converts by softmax, obtains every
The probability of a character, calculation formula are as follows:
Calculate loss function:
Wherein, ytIt is one-hot coding, represents the character of t moment output, P is output sequence length;
Step 2.3, it constructs attention module layer: obtaining the hiding shape at the last layer all moment in Encoder module layer
State vectorQ is the sum for inputting character;Hiding for first layer t-1 moment is obtained in Decoder module
State vectorThe calculation formula of attention vector is as follows:
α=softmax (w)
In calculating Decoder module layer when each layer of t moment of hidden state vector, β is added;
Step 3, carry out model training using Adam optimization method: after neural network model constructs successfully, data are from being input to
The whole flow process of output exists in the form of TensorFlow falls into a trap nomogram, directly utilizes tf.train.AdamOptimizer
() .minimize () is iterated training, finds parameter optimal value;
Step 4, error checking tasks are carried out, specifically:
In the Decoder stage, deduction process is carried out using beam search, beam size is 2, selects two probability every time most
High character when meeting full stop<EOS>, stops inferring, obtains output sequence as the input predicted next time.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811441697.0A CN109614612A (en) | 2018-11-29 | 2018-11-29 | A kind of Chinese text error correction method based on seq2seq+attention |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811441697.0A CN109614612A (en) | 2018-11-29 | 2018-11-29 | A kind of Chinese text error correction method based on seq2seq+attention |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109614612A true CN109614612A (en) | 2019-04-12 |
Family
ID=66005393
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811441697.0A Pending CN109614612A (en) | 2018-11-29 | 2018-11-29 | A kind of Chinese text error correction method based on seq2seq+attention |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109614612A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110188353A (en) * | 2019-05-28 | 2019-08-30 | 百度在线网络技术(北京)有限公司 | Text error correction method and device |
CN110705399A (en) * | 2019-09-19 | 2020-01-17 | 安徽七天教育科技有限公司 | Method for automatically identifying mathematical formula |
CN110866390A (en) * | 2019-10-15 | 2020-03-06 | 平安科技(深圳)有限公司 | Method and device for recognizing Chinese grammar error, computer equipment and storage medium |
CN111175852A (en) * | 2019-12-27 | 2020-05-19 | 中国电子科技集团公司第十四研究所 | Airport fog forecast early warning method based on long-time memory algorithm |
CN114548080A (en) * | 2022-04-24 | 2022-05-27 | 长沙市智为信息技术有限公司 | Chinese wrong character correction method and system based on word segmentation enhancement |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080059869A1 (en) * | 2006-09-01 | 2008-03-06 | The Regents Of The University Of California | Low cost, high performance error detection and correction |
CN107769972A (en) * | 2017-10-25 | 2018-03-06 | 武汉大学 | A kind of power telecom network equipment fault Forecasting Methodology based on improved LSTM |
CN108170686A (en) * | 2017-12-29 | 2018-06-15 | 科大讯飞股份有限公司 | Text interpretation method and device |
CN108491372A (en) * | 2018-01-31 | 2018-09-04 | 华南理工大学 | A kind of Chinese word cutting method based on seq2seq models |
-
2018
- 2018-11-29 CN CN201811441697.0A patent/CN109614612A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080059869A1 (en) * | 2006-09-01 | 2008-03-06 | The Regents Of The University Of California | Low cost, high performance error detection and correction |
CN107769972A (en) * | 2017-10-25 | 2018-03-06 | 武汉大学 | A kind of power telecom network equipment fault Forecasting Methodology based on improved LSTM |
CN108170686A (en) * | 2017-12-29 | 2018-06-15 | 科大讯飞股份有限公司 | Text interpretation method and device |
CN108491372A (en) * | 2018-01-31 | 2018-09-04 | 华南理工大学 | A kind of Chinese word cutting method based on seq2seq models |
Non-Patent Citations (2)
Title |
---|
徐俊 等: "基于seq2seq模型的中文纠错任务", 《人民网》 * |
杨林: "基于循环神经网络的英文写作错误检测及纠正的研究与实现", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110188353A (en) * | 2019-05-28 | 2019-08-30 | 百度在线网络技术(北京)有限公司 | Text error correction method and device |
CN110705399A (en) * | 2019-09-19 | 2020-01-17 | 安徽七天教育科技有限公司 | Method for automatically identifying mathematical formula |
CN110866390A (en) * | 2019-10-15 | 2020-03-06 | 平安科技(深圳)有限公司 | Method and device for recognizing Chinese grammar error, computer equipment and storage medium |
CN110866390B (en) * | 2019-10-15 | 2022-02-11 | 平安科技(深圳)有限公司 | Method and device for recognizing Chinese grammar error, computer equipment and storage medium |
CN111175852A (en) * | 2019-12-27 | 2020-05-19 | 中国电子科技集团公司第十四研究所 | Airport fog forecast early warning method based on long-time memory algorithm |
CN114548080A (en) * | 2022-04-24 | 2022-05-27 | 长沙市智为信息技术有限公司 | Chinese wrong character correction method and system based on word segmentation enhancement |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109614612A (en) | A kind of Chinese text error correction method based on seq2seq+attention | |
CN107203511B (en) | Network text named entity identification method based on neural network probability disambiguation | |
US20220147836A1 (en) | Method and device for text-enhanced knowledge graph joint representation learning | |
CN113239186B (en) | Graph convolution network relation extraction method based on multi-dependency relation representation mechanism | |
CN106682089A (en) | RNNs-based method for automatic safety checking of short message | |
US20220300546A1 (en) | Event extraction method, device and storage medium | |
CN105976056A (en) | Information extraction system based on bidirectional RNN | |
CN112035841B (en) | Intelligent contract vulnerability detection method based on expert rules and serialization modeling | |
CN109408574B (en) | Complaint responsibility confirmation system based on text mining technology | |
CN115131627B (en) | Construction and training method of lightweight plant disease and pest target detection model | |
CN109783644A (en) | A kind of cross-cutting emotional semantic classification system and method based on text representation study | |
CN111709244A (en) | Deep learning method for identifying causal relationship of contradictory dispute events | |
CN116089873A (en) | Model training method, data classification and classification method, device, equipment and medium | |
CN110362797A (en) | A kind of research report generation method and relevant device | |
CN108920446A (en) | A kind of processing method of Engineering document | |
CN115688920A (en) | Knowledge extraction method, model training method, device, equipment and medium | |
CN116484024A (en) | Multi-level knowledge base construction method based on knowledge graph | |
CN111597816A (en) | Self-attention named entity recognition method, device, equipment and storage medium | |
CN115062617A (en) | Task processing method, device, equipment and medium based on prompt learning | |
US20230401390A1 (en) | Automatic concrete dam defect image description generation method based on graph attention network | |
CN112559741A (en) | Nuclear power equipment defect recording text classification method, system, medium and electronic equipment | |
Li et al. | Evaluating BERT on cloud-edge time series forecasting and sentiment analysis via prompt learning | |
CN115827865A (en) | Method and system for classifying objectionable texts by fusing multi-feature map attention mechanism | |
CN113626537B (en) | Knowledge graph construction-oriented entity relation extraction method and system | |
CN115587192A (en) | Relationship information extraction method, device and computer readable 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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190412 |