CN115619192A - Hybrid relation extraction algorithm for demand planning rules - Google Patents
Hybrid relation extraction algorithm for demand planning rules Download PDFInfo
- Publication number
- CN115619192A CN115619192A CN202211408137.1A CN202211408137A CN115619192A CN 115619192 A CN115619192 A CN 115619192A CN 202211408137 A CN202211408137 A CN 202211408137A CN 115619192 A CN115619192 A CN 115619192A
- Authority
- CN
- China
- Prior art keywords
- supervised
- relation
- semi
- learning
- mfop
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000605 extraction Methods 0.000 title claims abstract description 61
- 238000013439 planning Methods 0.000 title claims abstract description 24
- 238000000034 method Methods 0.000 claims abstract description 53
- 238000012549 training Methods 0.000 claims description 17
- 238000004458 analytical method Methods 0.000 claims description 7
- 238000011156 evaluation Methods 0.000 claims description 3
- 238000002372 labelling Methods 0.000 claims description 3
- 239000000203 mixture Substances 0.000 claims description 3
- 230000000694 effects Effects 0.000 claims description 2
- 238000012935 Averaging Methods 0.000 claims 1
- 238000012360 testing method Methods 0.000 claims 1
- 238000005516 engineering process Methods 0.000 abstract description 4
- 238000013473 artificial intelligence Methods 0.000 abstract description 3
- 238000007726 management method Methods 0.000 description 3
- 108091028051 Numt Proteins 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- PXFBZOLANLWPMH-UHFFFAOYSA-N 16-Epiaffinine Natural products C1C(C2=CC=CC=C2N2)=C2C(=O)CC2C(=CC)CN(C)C1C2CO PXFBZOLANLWPMH-UHFFFAOYSA-N 0.000 description 1
- 101150071100 CBY2 gene Proteins 0.000 description 1
- 238000012550 audit Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 239000013307 optical fiber Substances 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000012552 review Methods 0.000 description 1
Images
Classifications
-
- 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
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0631—Resource planning, allocation, distributing or scheduling for enterprises or organisations
- G06Q10/06315—Needs-based resource requirements planning or analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/205—Parsing
- G06F40/211—Syntactic parsing, e.g. based on context-free grammar [CFG] or unification grammars
-
- 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
- G06Q50/06—Energy or water supply
-
- 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
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Landscapes
- Business, Economics & Management (AREA)
- Engineering & Computer Science (AREA)
- Human Resources & Organizations (AREA)
- Economics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Strategic Management (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- General Health & Medical Sciences (AREA)
- Marketing (AREA)
- Entrepreneurship & Innovation (AREA)
- Public Health (AREA)
- Game Theory and Decision Science (AREA)
- Educational Administration (AREA)
- Development Economics (AREA)
- Operations Research (AREA)
- Water Supply & Treatment (AREA)
- Quality & Reliability (AREA)
- Primary Health Care (AREA)
- Artificial Intelligence (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computational Linguistics (AREA)
- General Engineering & Computer Science (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a hybrid relation extraction algorithm facing to requirement planning rules, which comprises 4 stages, namely a parameter pre-extraction stage, an autonomous learning stage, an active learning stage and an application deployment stage on a general data set. Based on the existing artificial intelligence technology and the relation extraction theory, the method can extract the relation of a large number of unknown demand planning rules on the basis of semi-supervised learning, and has the relation extraction capability under weak supervision. The problem that noise is possibly large due to different levels of written requirement plans of various people in extraction can be solved. After the relationship extraction, unreasonable places which may exist are prompted to human users.
Description
Technical Field
The invention belongs to the field of electric power, and relates to a hybrid relation extraction algorithm for demand planning rules.
Background
Demand planning is the basis of intelligent procurement as the source of material management. With the increasing lean requirements of power grids on plan management and the increasing timeliness requirements of plan submission in recent years, the problem of low efficiency and high possibility of errors exists in the traditional manual mode for auditing the demand plan, the business chain of the plan management work is long, the related levels are multiple, the variety of the demand plan is complicated, the rule system relates to numerous constraint conditions such as business coding, business calculation, space-time range and the like, and the plan full-time and review experts need to manually audit the rules one by one. Moreover, because the demand plan relates to a plurality of batches, the auditing points are many, the association relationship is complex, and the detailed condition is difficult to grasp accurately. Therefore, a new demand plan information intelligent technology is urgently needed, complex logics in the existing auditing main points are extracted into machine rules capable of standardizing auditing quickly and efficiently by using tools of machine learning, deep learning and data science, the efficiency and accuracy of demand plan auditing are improved, and a tamping foundation is implemented for efficient purchasing.
The relation extraction method based on deep learning mainly comprises two types of remote supervision and supervised learning, wherein the supervised learning uses an artificially labeled data set in the training process, and the remote supervision method automatically labels corpora by aligning a remote knowledge base. The performance of the supervised and remote supervised models depends on the labeling quality of the training set, and the remote supervised data set is labeled in a mode that the remote supervised data set contains a large amount of noise, so that how to reduce the influence of the noise on the models is a key research problem of remote supervised relationship extraction.
For the task of the demand planning rule, the problems to be solved are as follows:
1. for the relationship extraction of a large number of unknown demand planning rules, the relationship extraction capability under weak supervision is required;
2. during extraction, the levels of various personnel for writing requirement plans are different, so that the noise is likely to be relatively large;
3. after the relationships are extracted, the relationships need to be approved by a human user, and according to requirements, a prompt for suspicious relationships should be provided, that is, unreasonable places may exist in an unknown demand plan, and the unreasonable places need to be prompted to the human user by a machine.
For the development of demand planning rules, there may be resources:
1. some common relations can be learned through some existing common public data sets;
2. a plurality of algorithm schemes with various characteristics exist, a plurality of algorithms can be used for learning, and hardware resources can support a plurality of algorithms for comparative learning;
3. in the initial stage of use, the approval process of the user is also a relation auditing process of unknown samples, and a new sample can be provided;
4. for the requirement planning rule of a large group, a general template can be defined, and when important sentences of core contents of the general template are written, a user can fill semantic contents in the general template, so that the automatic determination of the sentence relation of the core contents is realized.
Therefore, it is easy to see that, in order to fully utilize the above resource conditions and complete the problems to be solved, a hybrid relation extraction algorithm can be designed 1, a large number of results of relation extraction are learned by using a semi-supervised scheme, and the results are given to full-supervised learning; 2. training a plurality of full-supervision schemes by using the existing general data set-based normally accumulated sample data, and then supervising the extraction result of the relation of the semi-supervision schemes by using the voting scheme; 3. controlling whether the accuracy is overlarge by using the angle of stability; 4. finally, if multiple learning schemes are used, especially if their deviations are known, the voting scheme can be used to provide the most deviated contents to the user as suspicious or unreasonable objects, prompting the user for analysis and approval. In a word, the algorithm is similar to a teacher-student learning scheme with multiple teachers and students in the whole life on a learning model, and the relationship between the teachers and the students is continuously converted.
Disclosure of Invention
In order to overcome the problems in the prior art, the invention aims to provide a mixed relation extraction algorithm facing to requirement planning rules, based on the existing artificial intelligence technology and relation extraction theory, on the basis of semi-supervised learning, the relation extraction of standard samples is realized, and then comparison is carried out, unreasonable contents which are considered to possibly exist are prompted to a user for approval, and the capacity of meta-learning is realized through active learning.
The purpose of the invention is realized by the following technical scheme:
a hybrid relation extraction algorithm facing to requirement planning rules is characterized by comprising the following steps:
step 1, parameter pre-extraction stage on general data set
Step 1-1, defining a general template, when writing important sentences of core content, a user can fill semantic content in the general template to automatically determine the sentence relationship of the core content, so that a small rule base R1 with an initially empty general relationship is provided, and the content can be obtained in the future through simple text recognition and position positioning instead of a complex relationship extraction scheme. In addition, the modular format requires that the length of the bar be controlled to obtain the distance length extracted from the remote supervised relationship under the present universal templatelfThe numerical value of (c).
Step 1-2, on general non-professional full-tag data sets DatasetG (1) to DatasetG (n), using 90% of data to enable 1 semi-supervised learning algorithm MH,mThe fully supervised learning algorithms MF (1) to MF (m) learn respectively, and models MHt and MFt (1) to MFt (m) are trained. They were then tested for performance on the remaining 10% of the samples, specifically in calculating the difference in relationship extraction between 1 semi-supervised and m fully supervised voting methods over all n datasets, one dataset alone each. The specific work is as follows:
statistics are made on the results of all relations extraction RsG (mi) (mi =1 to mimax) for the remaining 10% of the full label samples of the individual dataset, where mimax is the total number of all ternary relations, which results are collectively summarized from the results of all learning algorithms. The semi-supervised learning algorithm yields in part RsGh (mih) (mih =1 to mihmax). The part obtained by the fully supervised learning algorithm is RsGf (mif) (mif =1 to mifmax), intersections exist between the RsG (mi), and the times of respectively recognizing by semi-supervised learning and fully supervised learning are also included in the whole RsG (mi). The relational expression of the ternary method is known in the art, and the main contents are ABC composition, A and C are objects, and B is a relation, for example, nanjing (A) belongs to (B) Jiangsu (C). In the invention, each ternary method relationship result also comprises a found array Timde1 (RsG, deth1, detf1, realout), wherein RsG represents that the pre-training is on a universal data set; deth1 is the result tested by the semi-supervised learning algorithm, and is only 1 or 0; detf1 is the number of times tested by m fully supervised algorithms, with values from 0 to m. For example, the rm-th relation "Nanjing belongs to Jiangsu" Timde1 (RsG, 1,2, 1) indicates that the relation is extracted by the semi-supervised relation and 2 times by the fully supervised relation, and Realout indicates whether the relation is artificially marked as correct in the fully labeled sample, only 1 or 0,1 represents correct, and 0 represents incorrect.
The fully supervised voting method sets a variable, votenum, which ranges from 1 to m. For the models MFt (1) to MFt (m) of the full supervision method, the voting method is carried out on the results in 10% of samples, and in Rsf (mi), the number of votes is found by the models MFt (1) to MFt (m) of the full supervision method to be more than or equal to votenum, so that the relation is judged to be true (1) from the perspective of full supervision. For example, if m is 10,votenum is 2, the rm-th Timde1 (RsG, 1,2, 1) in Timde1 (RsG, deth1, detf1, realout) in the above RsG (mi) that the relation "south kyo belongs to jiangsu" becomes Timde2 (RsG, deth2, detf2, realout) = (RsG, 1), detf2 is now the same as deth1, deth2, and only 1 or 0 indicates a confirmed relation. 1 indicates that the results of voten threshold voting by m fully supervised models are confirmed, at which point in this relationship semi-supervision works as well as fully supervised, and they all agree with the results of the true artificial label Realout.
Using a history method, setting the votenum from 1 to m, then in a single data set, finding out which value the result proportion of detf2 and Realout equal to each other is the highest in all the results, and obtaining the votembest (ni) on n data sets, wherein ni =1 to n. Then, these voteumbest (ni) are averaged to obtain voteop, which is used as the voting threshold obtained by the post-voting method.
Using the above fully supervised voting method, the threshold is voteop, and the accuracy of voteop is obtained in n universal data sets, pre (MFop (1)) -Pre (MFop (n)). And then calculating the average value avg (MFop) and the variance var (MFop) of the two values for later use.
In addition, the accuracy performance of the semi-supervised method in each individual dataset, pre (Mht (1)) -. Pre (Mht (n)), is calculated. And then calculating the average value avg (Mht) and the variance var (Mht) of the average value avg (Mht) and the variance var (Mht) for later use.
Step 2, autonomous learning phase on professional but weakly labeled data sets
And 2-1, learning a data set DatasetS of the weak tags by using a semi-supervised learning algorithm MH. Extracting the distance length obtained in the step 1-1 under the universal template and the remote supervision relationshiplfThe value of (d) is provided to the MH as its parameter. Meanwhile, a small rule base R1 of the initially empty general relation in the step 1-1 is used as an evaluation index, the autonomously learned data of the small rule base R1 is required to contain all the small rule bases R1, and the model MHt2 is trained.
Step 2-2, the learned semi-supervised learning algorithm MHt2 obtains knowledge results of all three relation sets for DatasetS, the knowledge results are listed as new data samples DatasetSF (1) of the fully supervised learning by taking sentences as unit length, and the data samples DatasetG (1) to DatasetG (n) are provided for DatasetG (1) to DatasetG (n)mAnd (3) learning by using all supervised learning algorithms MF (1) -MF (m), and training respective models MFt2 (1) -MFt 2 (m).
Step 3, active learning stage
At the moment, the algorithm starts to work for the examination and approval personnel, and the stage is to submit the contents with the largest difference of the results of a plurality of learning methods to the examination and approval personnel and ask the examination and approval personnel to label the contents.
And 3-1, performing relational extraction on each new sample by using the semi-supervised model MHt2 and the fully-supervised model MFt2 (1) -MFt 2 (m) trained in the step 2 to obtain an Rsf2 (mj) relationship set, wherein mj = 1-mjmax, and the Rsf2 (mj) relationship set comprises a group of triple results similar to the step 1-2 and comprises a discovered array Time 1 (Rsf 2, deth1, detf1, realout). At this point, because there is no tag, realout is Null. At this moment, when defh1 is calculated to be 0, detf1 is larger than 7; and defh1 is 1, derf1 is less than 2, namely the difference between semi-supervision and full supervision is large. These conditions are provided to the approver, who is asked to collectively implement the annotation, namely the annotation Realout.
And 3-2, manually marking the content, namely the knowledge result with the ternary relation set, taking a sentence as the unit length, listing the content as a new data sample DatasetSF (2), adding the data sample DatasetSF (2) into the data set in the step 2-2, and training a model MHt2 by referring to the step 2-2.
3-3, referring to the step 2-2, analyzing the new sample in the step 3-1 by using MHt2, and giving all the fully supervised learning of the ternary relationship extracted from the new sample in the step 3-1, and training respective models MFt2 (1) -MFt 2 (m);
and 3-4, repeating the steps 3-1 to 3-3 until no case that the difference of the step 3-1 is large occurs when at least 3 continuous samples are analyzed. Since the knowledge samples are incremental, the number of such widely differing cases is theoretically ever shrinking. Meanwhile, if the user wants to increase the speed and reduce the time, the label-free three-element relation group multi-active increasing labels of the step 3-1 can be added, so that the learning sample is improved.
Step 4, application deployment phase
The work of the stage is to face a new task, not to actively require the examining and approving personnel to provide the label of a dispute position, but to passively accept the examination and approval result of each task, and provide the inconsistent content of the analysis result of the fully-supervised and semi-supervised models to the user to remind the user that an unreasonable place possibly appears, namely to complete the problem 3 in the background of the invention.
And 4-1, performing relational extraction on each new sample by using the semi-supervised model MHt2 and the fully-supervised model MFt2 (1) -MFt 2 (m) trained in the step 3 to obtain an Rsf3 (mk) relationship set, wherein mk = 1-mkmax, and the Rsf3 (mk) relationship set comprises a group of triple results similar to the step 1-2 and comprises a found array Timkdde 1 (Rsf 2, deth1, detf1, realout). At this point, because there is no tag, realout is Null. At this moment, referring to step 1-2, the voteop obtained in step 1-2 is recorded in this step, and when derf1 is greater than or equal to voteop, derf2 in timkdde 2 (Rsf 2, deth2, detf2, realout) is 1, otherwise, it is 0. All content parts with deth2 and detf2 inconsistent are displayed to the user, reminding him of possible unreasonable places to appear here.
And 4-2, when the user approves the sample in the step 4-1, obtaining a result of failing to pass (0) or passing (1), and taking the result as a weak label to be taken as a data set together with the data set in the step 2. And (3) performing centralized learning and updating every NumT1 samples according to the step 2 to obtain a new trained semi-supervised model MHt2 and fully supervised models MFt2 (1) -MFt 2 (m).
And 4-3, performing centralized learning statistics according to the step 1-2 every NumT2 (wherein NumT2> NumT 1) samples. And obtaining a new trained semi-supervised model MHt2' and a fully supervised model MFt2' (1) -MFt 2' (m) by using 90% of the data in the step 4-2. Counting the average value avg '(MFop) and the variance var' (MFop) of the fully supervised model; the accuracy performance of the semi-supervised method is counted, and the average value avg' (Mht) of the accuracy performance is calculated.
Condition 1: and analyzing whether the accurate range B1 of the average value of the fully supervised model in the step 3 is within the accurate range B0 of the 3 variances obtained in the step 1-2.
B0: (avg (MFop)-3*var(MFop)) ~ (avg(MFop)+3*var(MFop))
B1: (avg’(MFop)-var’(MFop) ) ~(avg’(MFop)+var’(MFop))
Condition 2: and analyzing whether the avg' (Mht) of the semi-supervised model mean value in the step 3 is within the 3 variance accurate range C0 obtained in the step 1-2.
C0: (avg(Mht)- 3*var’(Mht)) ~ (avg(Mht) + 3*var’(Mht))
If the condition 1 or the condition 2 is not one, the excessive noise is learned, and the step 3 needs to be re-entered, and the deployment is performed after the accurate label is learned.
The model mainly comprises a semi-supervised learning algorithm MH and a plurality of fully supervised learning algorithms MF (1) -MF (m). The data set comprises a plurality of universal but non-professional full label data sets DatasetG (1) -DatasetG (n), and a sample set DatasetS of a professional but only requirement planning rule whether the whole is correct or not, wherein the sample of the sample set is a sample added with a weak label in each approval. In the learning process of the invention, the invention comprises 4 stages of parameter pre-extraction stage, autonomous learning stage, active learning stage and application deployment stage on a general data set. In addition, the invention can also define a general template, when writing the important sentence of the core content, the user can fill the semantic content in, through simple semantic recognition, according to the position in the text, can realize the automatic determination of the sentence relation of the core content, thus having a small rule base R1 of the general relation.
The invention has the beneficial effects that:
by implementing the scheme, 3 problems of relation extraction facing to the requirement planning rule can be solved. The method can extract the relations of a large number of unknown demand planning rules, and has the capacity of extracting the relations under weak supervision. The problem that noise is possibly large due to different levels of written requirement plans of various people in extraction can be solved. After the relationship extraction, unreasonable places which may exist are prompted to human users.
Drawings
FIG. 1 is a flowchart of the overall algorithm of an embodiment of the present invention.
DETAILED DESCRIPTION OF EMBODIMENT (S) OF INVENTION
The invention is further illustrated with reference to fig. 1.
The invention provides an algorithm for extracting the relation in the requirement planning rule of the large-scale enterprise based on the existing artificial intelligence technology and the relation extraction theory, which realizes the relation extraction of standard samples on the basis of semi-supervised learning, then compares the relation extraction and the comparison, prompts the contents which are considered to be unreasonable to the user for approval, and realizes the meta-learning ability through active learning.
The model mainly comprises a semi-supervised learning algorithm MH and a plurality of fully supervised learning algorithms MF (1) -MF (m). The data set comprises a plurality of universal but non-professional full label data sets DatasetG (1) -DatasetG (n), and a sample set DatasetS of a professional but only requirement planning rule whether the whole is correct or not, wherein the sample of the sample set is added with a weak label sample in each approval. In the learning process, the method comprises 4 stages of parameter pre-extraction, autonomous learning, active learning and application deployment on a universal data set. In addition, the invention can also define a general template, when writing the important sentence of the core content, the user can fill the semantic content in, and realize the automatic determination of the sentence relation of the core content, thus having a small rule base R1 of the general relation.
In a specific embodiment of the invention, the semi-supervised learning algorithm MH uses the TMNN algorithm in the literature (nijun. Study of relationship extraction methods based on weakly supervised learning [ D ]. University of big graduate.). A total of 10 learning algorithms, i.e. m =10, are:
[1] multi-head, which treats the relationship extraction task as a Multi-head selection problem, capable of extracting multiple relationship types between pairs of entities.
[2] Multi-head + AT, the model performs entity relationship extraction using a Multi-head attention mechanism based confrontational learning method.
[3] Sci IE, the model introduces a multitasking setup for classifying entities, relationships and co-occurring word clusters in scientific articles. The relational classification model can utilize the relation of cross sentences and reduce cascade errors through co-occurrence word connection.
[4] relationship-Metric, model combines Metric learning and convolutional neural networks to achieve relationship extraction.
[5] Biaffine Attention, which extends the Bi LSTM-CRF model to learn the second-order interaction of hidden states using a Deep double affine Attention network Layer (Deep Biaffine At-extension Layer).
[6] Multi-turn QA, the model defines entity relationship extraction as a Multi-turn dialogue question-answer task, which achieves SOTA (State-of-the-art) effect in the sequence annotation method.
[7] Dy GIE + +, further expanding the span-based Dy GIE model, and introducing co-fingering analysis to realize the representation of enhanced entities and relationship characteristics.
[8] SpERT is a span attention-based entity-relationship extraction method that achieves the best results on Co NLL04 and ADE datasets.
[9] The Hierarchical Attention mechanism is adopted, and the most important semantic information can be captured to enhance the relation extraction capability.
[10] Cas Rel, which first identifies all possible head entities and second uses a specific relationship tagger to identify the corresponding relationship of each head entity and tail entity.
In this embodiment, the general but non-professional full label dataset is a TACRED dataset, a sim Eval-2010 Task 8 dataset, a SCIERC dataset, a CoNLL04 dataset, and an ADE dataset, that is, n =5. The sample set DatasetS of the requirement planning rule that is professional but is only wholly correct or not is self-created.
The method comprises the following steps:
step 1, parameter pre-extraction stage on universal data set
Step 1-1, defining a general template, when writing important sentences of core content, a user can fill semantic content in the general template to automatically determine the sentence relationship of the core content, so that a small rule base R1 with an initially empty general relationship is provided, and the content can be obtained in the future through simple text recognition and position positioning instead of a complex relationship extraction scheme. In addition, the module format requirement is set, the lengths of the sections can be controlled, and therefore the distance length extracted by the remote supervision relation under the universal template is obtainedlfThe numerical value of (c).
In this embodiment, R1 is self-created, and has 78 standard ternary semantic rules, including multiple categories such as a subunit of the first party, a financial relationship, a task relationship, and the like, and as long as the accurate position is filled in, the 78 standard ternary semantic rules, rather than the 78 standard ternary semantic rules, can be automatically analyzed through text recognitionIs extracted. In addition, since the standard length is established, it is possible to set the length of the optical fiber to a predetermined valuelfIs limited to 80 Chinese characters.
Step 1-2, on general non-professional full-label data sets DatasetG (1) to DatasetG (n), using 90% of data to enable 1 semi-supervised learning algorithm MH,mThe fully supervised learning algorithms MF (1) to MF (m) learn respectively, and models MHt and MFt (1) to MFt (m) are trained. They were then tested for performance on the remaining 10% of the samples, specifically in calculating the difference in relationship extraction between 1 semi-supervised and m fully supervised voting methods over all n datasets, one dataset alone each. In the present embodiment, m =10,n =5. The specific work is as follows:
statistics are made on the results of all relations extraction RsG (mi) (mi =1 to mimax) for the remaining 10% of the full label samples of the individual dataset, where mimax is the total number of all ternary relations, which results are collectively summarized from the results of all learning algorithms. The semi-supervised learning algorithm yields in part RsGh (mih) (mih =1 to mihmax). The part obtained by the fully supervised learning algorithm is RsGf (mif) (mif =1 to mifmax), intersections exist between the RsG (mif), and the times respectively recognized by semi-supervised learning and fully supervised learning are also included in the whole RsG (mi). The relational expression of the ternary method is known in the art, and the main contents of the ternary method are ABC composition, A and C are objects, and B is a relation, for example, nanjing (A) belongs to (B) Jiangsu (C). In the invention, each ternary method relationship result also comprises a found array Timde1 (RsG, deth1, detf1, realout), wherein RsG represents that the pre-training is on a universal data set; deth1 is the result tested by the semi-supervised learning algorithm, and is only 1 or 0; detf1 is the number of times tested by m fully supervised algorithms, with values from 0 to m. For example, the rm relationship "Nanjing belongs to Jiangsu" is Timde1 (RsG, 1,2, 1), which indicates that the relationship is extracted by the semi-supervised relationship and 2 times by the fully supervised relationship, and Realout is whether the relationship is manually marked as correct in the fully labeled sample, only 1 or 0,1 represents correct, and 0 represents incorrect.
The fully supervised voting method sets a variable, votenum, which ranges from 1 to m. For the models MFt (1) to MFt (m) of the full supervision method, the voting method is carried out on the results in 10% of samples, and in Rsf (mi), the number of votes is found by the models MFt (1) to MFt (m) of the full supervision method to be more than or equal to votenum, so that the relation is judged to be true (1) from the perspective of full supervision. For example, if m is 10,votenum is 2, the rm-th relation "mimde 1 (RsG, 1,2, 1) of" nanjing belongs to jiangsu "in Timde1 (RsG, deth1, detf1, realout) in the above RsG (mi) will become Timde2 (RsG, deth2, detf2, realout) = (RsG, 1), and detf2 is now the same as deth1, deth2, and is only 1 or 0, indicating a confirmed relation. 1 indicates that the results of voten threshold voting by m fully supervised models are confirmed, at which point in this relationship semi-supervision works as well as fully supervised, and they all agree with the results of the true artificial label Realout.
Using a history method, setting the votenum from 1 to m, then in a single data set, finding out which value the result proportion of detf2 and Realout equal to each other is the highest in all the results, and obtaining the votembest (ni) on n data sets, wherein ni =1 to n. Then, these votembubest (ni) are averaged to obtain voteop as a voting threshold obtained by a post-voting method.
In this specific embodiment, voteop =3.
Using the above fully supervised voting method, the threshold is voteop, and the accuracy of voteop is obtained in n universal data sets, pre (MFop (1)) -Pre (MFop (n)). And then calculating the average value avg (MFop) and the variance var (MFop) of the two values for later use.
In addition, the accuracy performance of the semi-supervised method, pre (Mht (1)) -Pre (Mht (n)), in each individual dataset is calculated. And then calculating the average value avg (Mht) and the variance var (Mht) of the average value avg (Mht) and the variance var (Mht) for later use.
Step 2, autonomous learning phase on professional but weakly labeled data sets
And 2-1, learning a data set DatasetS of the weak label by using a semi-supervised learning algorithm MH. Extracting the distance length obtained in the step 1-1 under the universal template and the remote supervision relationshiplfIs provided to the MH as its parameter. Meanwhile, a channel in step 1-1 is initially emptyAnd (3) taking the small rule base R1 of the relation as an evaluation index, and requiring the autonomously learned data to contain all the small rule bases R1 and training the MHt2.
In this embodiment, R1 is extracted through text semantic autonomous analysis, not relation extraction, and projects semantics into the known 78 rules according to the position at the time of writing, so that the analysis can be performed simply. And then learn by providing the MH with these 78 rules.
Step 2-2, the learned half-supervised learning algorithm MHt2 is used for obtaining knowledge results of all three relation sets for DatasetS, sentences are taken as unit length, the knowledge results are listed as new data samples DatasetSF (1) for full-supervised learning, and the data samples DatasetG (1) -DatasetG (n) are provided for the DatasetG (1) -DatasetG (n)mAnd (3) learning by using all supervised learning algorithms MF (1) -MF (m), and training respective models MFt2 (1) -MFt 2 (m).
Step 3, active learning phase
At the moment, the algorithm starts to work for the examination and approval personnel, and the stage is to submit the contents with the largest difference of the results of a plurality of learning methods to the examination and approval personnel and ask the examination and approval personnel to label the contents.
And 3-1, performing relational extraction on each new sample by using the semi-supervised model MHt2 and the fully-supervised models MFt2 (1) -MFt 2 (m) trained in the step 2 to obtain an Rsf2 (mj) relationship set, wherein mj = 1-mjmax, and the Rsf2 (mj) relationship set comprises a group of triple results similar to the step 1-2 and comprises a discovered array Time 1 (Rsf 2, deth1, detf1, realout). At this point, because there is no tag, realout is Null. At this moment, when defh1 is calculated to be 0, detf1 is larger than 7; and defh1 is 1, derf1 is less than 2, namely the difference between semi-supervision and full supervision is large. These conditions are provided to the approver, who is asked to collectively implement the annotation, namely the annotation Realout.
And 3-2, manually marking the content, namely the knowledge result with the ternary relation set, taking a sentence as the unit length, listing the content as a new data sample DatasetSF (2), adding the data sample DatasetSF (2) into the data set in the step 2-2, and training a model MHt2 by referring to the step 2-1.
3-3, referring to the step 2-2, analyzing the new sample in the step 3-1 by using MHt2, and giving all the fully supervised learning of the ternary relationship extracted from the new sample in the step 3-1, and training respective models MFt2 (1) -MFt 2 (m);
and 3-4, repeating the steps 3-1 to 3-3 until no case that the difference of the step 3-1 is large occurs when at least 3 continuous samples are analyzed. Since the knowledge samples are incremental, the number of such widely different cases is theoretically shrinking. Meanwhile, if the user wants to increase the speed and reduce the time, the label-free three-element relation group multi-active increasing labels of the step 3-1 can be added, so that the learning sample is improved.
Step 4, application deployment phase
The work of the stage is to face a new task, not to actively require the examining and approving personnel to provide the label of a dispute position, but to passively accept the examination and approval result of each task, and provide the inconsistent content of the analysis result of the fully-supervised and semi-supervised models to the user to remind the user that an unreasonable place possibly appears, namely to complete the problem 3 in the background of the invention.
And 4-1, performing relational extraction on each new sample by using the semi-supervised model MHt2 and the fully-supervised model MFt2 (1) -MFt 2 (m) trained in the step 3 to obtain an Rsf3 (mk) relationship set, wherein mk = 1-mkmax, and the Rsf3 (mk) relationship set comprises a group of triple results similar to the step 1-2 and comprises a found array Timkdde 1 (Rsf 2, deth1, detf1, realout). At this point, because there is no tag, realout is Null. At this moment, referring to step 1-2, the voteop obtained in step 1-2 is recorded in this step, and when derf1 is greater than or equal to voteop, derf2 in timkdde 2 (Rsf 2, deth2, detf2, realout) is 1, otherwise, it is 0. All content parts with deth2 and detf2 inconsistent are displayed to the user, reminding the user that unreasonable places may appear here.
And 4-2, when the user approves the sample in the step 4-1, obtaining a result of failing to pass (0) or passing (1), and taking the result as a weak label to be taken as a data set together with the data set in the step 2. And (3) performing centralized learning and updating every NumT1 samples according to the step 2 to obtain a new trained semi-supervised model MHt2 and fully supervised models MFt2 (1) -MFt 2 (m). In this embodiment, numT1=100.
And 4-3, performing centralized learning statistics according to the step 1-2 every NumT2 (wherein NumT2> NumT 1) samples. And obtaining a new trained semi-supervised model MHt2' and a new trained fully supervised model MFt2' (1) -MFt 2' (m) by using 90% of the data in the step 4-2. Counting the average value avg '(MFop) and the variance var' (MFop) of the fully supervised model; the accuracy performance of the semi-supervised method is counted, and the average value avg' (Mht) of the accuracy performance is calculated. In this embodiment, numT1=300.
Condition 1: and analyzing whether the accurate range B1 of the average value of the existing fully supervised model in the step 3 is within 3 accurate ranges B0 of the variance obtained in the step 1-2.
B0: (avg (MFop)-3*var(MFop)) ~ (avg(MFop)+3*var(MFop))
B1: (avg’(MFop)-var’(MFop) ) ~(avg’(MFop)+var’(MFop))
Condition 2: and analyzing whether the avg' (Mht) of the semi-supervised model mean value in the step 3 is within the 3 variance accurate range C0 obtained in the step 1-2.
C0: (avg(Mht)- 3*var’(Mht)) ~ (avg(Mht) + 3*var’(Mht))
If the condition 1 or the condition 2 is not one, the excessive noise is learned, and the step 3 needs to be re-entered, and the deployment is performed after the accurate label is learned.
By implementing the scheme, 3 problems of relation extraction facing to the requirement planning rule can be solved. The method can extract the relations of a large number of unknown demand planning rules, and has the capacity of extracting the relations under weak supervision. The problem that noise is possibly large due to different levels of written requirement plans of various people in extraction can be solved. After the relationship extraction, the unreasonable place possibly existing is prompted to a human user.
Claims (4)
1. A hybrid relation extraction algorithm for demand planning rules is characterized by comprising the following steps:
step 1, parameter pre-extraction stage on general data set
Step 1-1, setting format requirements of the universal template, and controlling the length of the subsection so as to obtain the distance length extracted by the remote supervision relation under the universal templatelfThe value of (d);
step 1-2, on a general non-professional full-label data set DatasetG (1) -DatasetG (n), using 90% of data as a training set, using the remaining 10% as a test set, obtaining a voting threshold voteop, and the average value and variance of the accuracy of the voting threshold voteop on n general data sets, and determining a stable range for later use;
step 2, autonomous learning phase on professional but weakly labeled data sets
Step 2-1, learning a data set DatasetS of the weak tags by using a semi-supervised learning algorithm MH; extracting distance length of remote supervision relation under a universal template obtained in the step 1-1lfThe value of (d) is provided to the MH as its parameter; meanwhile, a small rule base R1 of the initially empty general relation in the step 1-1 is used as an evaluation index, the autonomously learned data of the small rule base R1 is required to contain all the small rule bases R1, and a semi-supervised model MHt2 is trained;
step 2-2, the learned half-supervised learning algorithm MHt2 is used for obtaining knowledge results of all three relation sets for DatasetS, sentences are taken as unit length, the knowledge results are listed as new data samples DatasetSF (1) for full-supervised learning, and the data samples DatasetG (1) -DatasetG (n) are provided for the DatasetG (1) -DatasetG (n)mLearning by using fully supervised learning algorithms MF (1) -MF (m), and training respective models MFt2 (1) -MFt 2 (m);
step 3, active learning phase
Step 3-1, performing relation extraction on each new sample by using the semi-supervised model MHt2 and the fully supervised models MFt2 (1) -MFt 2 (m) trained in the step 2 to obtain an Rsf2 (mj) relation set, wherein mj = 1-mjmax, and the Rsf2 (mj) relation set comprises a group of triple results similar to the step 1-2 and comprises a discovered array Time 1 (Rsf 2, deth1, detf1, realout); at this moment, because no label exists, realout is Null; at this moment, when defh1 is calculated to be 0, detf1 is larger than 7; and defh1 is 1, derf1 is less than 2, namely the difference between the semi-supervision and the full supervision is large; labeling the situations collectively, namely labeling Realout;
step 3-2, manually marking the content, namely the knowledge result with the ternary relation set, taking a sentence as the unit length, listing the content as a new data sample DatasetSF (2), adding the data sample DatasetSF (2) into the data set in the step 2-2, and training a semi-supervised model MHt2 with reference to the step 2-2;
3-3, referring to the step 2-2, analyzing the new sample in the step 3-1 by using MHt2, and giving all the fully supervised learning of the ternary relationship extracted from the new sample in the step 3-1, and training respective models MFt2 (1) -MFt 2 (m);
step 3-4, repeating the step 3-1 to the step 3-3 until no case that the difference of the step 3-1 is large occurs when at least 3 continuous samples are analyzed;
step 4, application deployment phase
Step 4-1, extracting the relation of each new sample by using the semi-supervised model MHt2 and the fully supervised model MFt2 (1) -MFt 2 (m) trained in the step 3 to obtain an Rsf3 (mk) relationship set, wherein mk = 1-mkmax, which contains a group of triple results similar to the step 1-2 and comprises a found array Timkdde 1 (Rsf 2, deth1, detf1, realout); at the moment, because no label exists, the Realout is Null, at the moment, referring to the step 1-2, the voteop obtained in the step 1-2 is recorded in the step, when derf1 is more than or equal to the voteop, derf2 in Timkdde 2 (Rsf 2, deth2, detf2, realout) is 1, otherwise, the derf2 is 0; displaying all content parts with inconsistent deth2 and detf2 to a user to remind the user of possible unreasonable places;
step 4-2, obtaining a result which does not pass (0) or passes (1) after the user approves the sample in the step 4-1, and taking the result as a weak label which is taken as a data set together with the data set in the step 2; every other NumT1 samples, learning and updating are carried out in a centralized mode according to the step 2, and new trained semi-supervised models MHt2 and fully supervised models MFt2 (1) -MFt 2 (m) are obtained;
4-3, performing centralized learning statistics on every other NumT2 samples, wherein NumT2 is greater than NumT1 according to the step 1-2; obtaining a new trained semi-supervised model MHt2' and a new trained fully supervised model MFt2' (1) -MFt 2' (m) by using 90% of the data in the step 4-2; counting the average value avg '(MFop) and the variance var' (MFop) of the fully supervised model; counting the accuracy performance of the semi-supervised method, and calculating the average value avg' (Mht) of the accuracy performance;
condition 1: analyzing whether the accurate range B1 of the average value of the fully supervised model in the step 3 is within the accurate range B0 of the 3 variances obtained in the step 1-2;
B0: (avg (MFop)-3*var(MFop)) ~ (avg(MFop)+3*var(MFop))
B1: (avg’(MFop)-var’(MFop) ) ~(avg’(MFop)+var’(MFop))
condition 2: analyzing whether the avg' (Mht) of the semi-supervised model mean value in the step 3 is within the 3 variance accurate range C0 obtained in the step 1-2;
C0: (avg(Mht)- 3*var’(Mht)) ~ (avg(Mht) + 3*var’(Mht))
if the condition 1 or the condition 2 is not one, the excessive noise is learned, and the step 3 needs to be re-entered, and the deployment is performed after the accurate label is learned.
2. The demand planning rule-oriented hybrid relationship extraction algorithm of claim 1, wherein the step 1 is specifically as follows:
step 1-1, writing important sentences of core contents of a universal template by defining the universal template, realizing automatic determination of the sentence relation of the core contents, and having a small rule base R1 of the initially empty universal relation, wherein the contents of the small rule base R1 can be obtained in the future through simple text recognition and position positioning instead of a complex relation extraction scheme; setting the format requirement of the module, controlling the length of the subsection so as to obtain the distance length extracted by the remote supervision relation under the universal templatelfThe value of (d);
step 1-2, on general non-professional full-label data sets DatasetG (1) to DatasetG (n), using 90% of data to enable 1 semi-supervised learning algorithm MH,mRespectively learning by using fully supervised learning algorithms MF (1) -MF (m), and training respective models MHt and MFt (1) -MFt (m); they were then tested for performance on the remaining 10% of the samples, calculated asThe difference in relation extraction between 1 semi-supervised and m fully supervised voting methods on each of these n data sets, one data set alone.
3. The demand planning rule-oriented hybrid relationship extraction algorithm of claim 2, wherein the steps 1-2 specifically work as follows:
counting the results of all relations extraction RsG (mi) (mi =1 to mimax) of the remaining 10% of the full label samples of the individual dataset, wherein mimax is the total number of all ternary relations, and the results are obtained by collectively collecting the results of all learning algorithms; the semi-supervised learning algorithm results in part in RsGh (mih) (mih =1 to mihmax); the part obtained by the fully supervised learning algorithm is RsGf (mif) (mif =1 to mifmax), intersections exist among the RsG (mi), and the times respectively recognized by semi-supervised learning and fully supervised learning are also included in the whole RsG (mi); the relational expression of the ternary method is known in the art, the main content of the ternary method is ABC composition, A and C are objects, B is a relation, each ternary method relation result further comprises a discovered array Timde1 (RsG, deth1, detf1, realout), wherein RsG represents that the pretraining is on a universal data set; deth1 is the result tested by the semi-supervised learning algorithm, and is only 1 or 0; detf1 is the number of times tested by m fully supervised algorithms, and has a value from 0 to m;
the fully supervised voting method sets a variable votenum, which ranges from 1 to m; performing a voting method on results of the models MFt (1) -MFt (m) of the full supervision method in 10% of samples, and judging that the relation is kept to be true (1) from the perspective of full supervision if the models MFt (1) -MFt (m) of the full supervision method in Rsf (mi) are found to be more than or equal to votenum times; 1 represents that the voting result of the voteum threshold by m fully supervised models is confirmed, at this moment, the semi-supervision and the fully supervised models have the same effect on the relation, and all of them accord with the result of the real artificial label Realout;
setting the votenum from 1 to m by using a traversal method, then in a single data set, finding out the value under which the equivalent result proportion of detf2 and Realout is the highest in all results, and obtaining the votenust (ni) on n data sets, wherein ni =1 to n;
then averaging the votembubest (ni) to obtain voteop which is used as a voting threshold value obtained by a later voting method;
using the above full-supervision voting method, wherein the threshold is voteop, and obtaining the accuracy of the voteop on n universal data sets Pre (MFop (1)) -Pre (MFop (n)); then, calculating the average value avg (MFop) and the variance var (MFop) of the two values for later use; in addition, calculating the accuracy performance of a semi-supervised method in each individual data set, namely Pre (Mht (1)) -Pre (Mht (n)); and then calculating the average value avg (Mht) and the variance var (Mht) of the average value avg (Mht) and the variance var (Mht) for later use.
4. The hybrid relationship extraction algorithm for demand planning rules according to claim 1, wherein in step 4, the inconsistent analysis results of the fully supervised and semi supervised models are provided to the user to remind the user of the possible unreasonable places.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211408137.1A CN115619192B (en) | 2022-11-10 | 2022-11-10 | Mixed relation extraction method oriented to demand planning rules |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211408137.1A CN115619192B (en) | 2022-11-10 | 2022-11-10 | Mixed relation extraction method oriented to demand planning rules |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115619192A true CN115619192A (en) | 2023-01-17 |
| CN115619192B CN115619192B (en) | 2023-10-03 |
Family
ID=84877808
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211408137.1A Active CN115619192B (en) | 2022-11-10 | 2022-11-10 | Mixed relation extraction method oriented to demand planning rules |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115619192B (en) |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106294593A (en) * | 2016-07-28 | 2017-01-04 | 浙江大学 | In conjunction with subordinate clause level remote supervisory and the Relation extraction method of semi-supervised integrated study |
| US20180137435A1 (en) * | 2016-11-17 | 2018-05-17 | Kabushiki Kaisha Toshiba | Information extraction support device, information extraction support method and computer program product |
| CN110427629A (en) * | 2019-08-13 | 2019-11-08 | 苏州思必驰信息科技有限公司 | Semi-supervised text simplified model training method and system |
| CN111737497A (en) * | 2020-06-30 | 2020-10-02 | 大连理工大学 | Weak supervision relation extraction method based on multi-source semantic representation fusion |
| CN112906368A (en) * | 2021-02-19 | 2021-06-04 | 北京百度网讯科技有限公司 | Industry text increment method, related device and computer program product |
| CN113627190A (en) * | 2021-08-27 | 2021-11-09 | 上海复深蓝软件股份有限公司 | Visualized data conversion method and device, computer equipment and storage medium |
| CN113936217A (en) * | 2021-10-25 | 2022-01-14 | 华中师范大学 | Priori semantic knowledge guided high-resolution remote sensing image weakly supervised building change detection method |
| CN114884894A (en) * | 2022-04-18 | 2022-08-09 | 南京邮电大学 | Semi-supervised network traffic classification method based on transfer learning |
-
2022
- 2022-11-10 CN CN202211408137.1A patent/CN115619192B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106294593A (en) * | 2016-07-28 | 2017-01-04 | 浙江大学 | In conjunction with subordinate clause level remote supervisory and the Relation extraction method of semi-supervised integrated study |
| US20180137435A1 (en) * | 2016-11-17 | 2018-05-17 | Kabushiki Kaisha Toshiba | Information extraction support device, information extraction support method and computer program product |
| CN110427629A (en) * | 2019-08-13 | 2019-11-08 | 苏州思必驰信息科技有限公司 | Semi-supervised text simplified model training method and system |
| CN111737497A (en) * | 2020-06-30 | 2020-10-02 | 大连理工大学 | Weak supervision relation extraction method based on multi-source semantic representation fusion |
| CN112906368A (en) * | 2021-02-19 | 2021-06-04 | 北京百度网讯科技有限公司 | Industry text increment method, related device and computer program product |
| CN113627190A (en) * | 2021-08-27 | 2021-11-09 | 上海复深蓝软件股份有限公司 | Visualized data conversion method and device, computer equipment and storage medium |
| CN113936217A (en) * | 2021-10-25 | 2022-01-14 | 华中师范大学 | Priori semantic knowledge guided high-resolution remote sensing image weakly supervised building change detection method |
| CN114884894A (en) * | 2022-04-18 | 2022-08-09 | 南京邮电大学 | Semi-supervised network traffic classification method based on transfer learning |
Non-Patent Citations (4)
| Title |
|---|
| QI YAO 等: "Saliency Guided Self-Attention Network for Weakly and Semi-Supervised Semantic Segmentation", 《IEEE ACCESS》, vol. 8, pages 14413 - 14423, XP011768413, DOI: 10.1109/ACCESS.2020.2966647 * |
| 姚祺: "基于自注意力机制的弱监督和半监督语义分割", 《中国优秀硕士学位论文全文数据库 信息科技辑》, no. 2, pages 138 - 2804 * |
| 彭博: "基于ALBERT的网络文物信息资源实体关系抽取方法研究", 《情报杂志》, vol. 41, no. 8, pages 156 - 162 * |
| 罗耀东: "湿地实体识别与开放关系抽取的研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》, no. 1, pages 138 - 2703 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115619192B (en) | 2023-10-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108717433A (en) | A kind of construction of knowledge base method and device of programming-oriented field question answering system | |
| CN110888989B (en) | Intelligent learning platform and construction method thereof | |
| CN116645129A (en) | Manufacturing resource recommendation method based on knowledge graph | |
| CN116484024A (en) | Multi-level knowledge base construction method based on knowledge graph | |
| CN116561264A (en) | Knowledge graph-based intelligent question-answering system construction method | |
| CN111666374A (en) | Method for integrating additional knowledge information into deep language model | |
| CN111241153A (en) | Enterprise natural person entity comprehensive judgment alignment method and system | |
| CN116342167B (en) | Intelligent cost measurement method and device based on sequence labeling named entity recognition | |
| Lou | [Retracted] Evaluation of College English Teaching Quality Based on Improved BT‐SVM Algorithm | |
| CN117033603A (en) | Construction method, device, equipment and storage medium of large model in vertical field | |
| CN115619192B (en) | Mixed relation extraction method oriented to demand planning rules | |
| CN113610194B (en) | Automatic classification method for digital files | |
| CN113222471B (en) | Asset wind control method and device based on new media data | |
| CN114519092A (en) | Large-scale complex relation data set construction framework oriented to Chinese field | |
| CN114281998A (en) | Multi-level annotator-oriented event annotation system construction method based on crowdsourcing technology | |
| Yu et al. | Data mining technology in the analysis of college students’ psychological problems | |
| He et al. | A credible predictive model for employment of college graduates based on LightGBM | |
| CN114510649B (en) | Social network and LSTM model accuracy calculating method based on deduplication sample | |
| CN117934209B (en) | Regional power system carbon emission big data analysis method based on knowledge graph | |
| CN117540727B (en) | Subjective question scoring method and system based on ALBERT model and RPA technology | |
| Feifei et al. | Intelligent question and answer analysis model of power ICT based on BI-LSTM-CRF | |
| CN113158079B (en) | Case public opinion timeline generation method based on difference case elements | |
| Jia | [Retracted] Random Algorithm and Skill Evaluation System Based on the Combing of Construction Mechanism of Higher Vocational Professional Group | |
| Arandi et al. | Bidirectional recommendation in HR analytics through text summarization | |
| He et al. | An Extensible Heterogeneous Network Embedding Framework for Knowledge Tracing. |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |