CN110781316A - Time perception knowledge representation learning method integrating hyperplane and duration modeling - Google Patents
Time perception knowledge representation learning method integrating hyperplane and duration modeling Download PDFInfo
- Publication number
- CN110781316A CN110781316A CN201911020394.6A CN201911020394A CN110781316A CN 110781316 A CN110781316 A CN 110781316A CN 201911020394 A CN201911020394 A CN 201911020394A CN 110781316 A CN110781316 A CN 110781316A
- Authority
- CN
- China
- Prior art keywords
- knowledge
- time
- duration
- effective
- hyperplane
- 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
- 238000000034 method Methods 0.000 title claims abstract description 75
- 230000008447 perception Effects 0.000 title claims abstract description 8
- 238000012549 training Methods 0.000 claims abstract description 22
- 238000004364 calculation method Methods 0.000 claims abstract description 11
- 230000006870 function Effects 0.000 claims description 25
- 239000013598 vector Substances 0.000 claims description 25
- 238000011156 evaluation Methods 0.000 claims description 19
- 230000002085 persistent effect Effects 0.000 claims description 11
- 230000001052 transient effect Effects 0.000 claims description 7
- 238000005070 sampling Methods 0.000 claims description 5
- 230000002688 persistence Effects 0.000 claims description 3
- 230000002123 temporal effect Effects 0.000 claims description 3
- 230000001186 cumulative effect Effects 0.000 claims description 2
- 238000005315 distribution function Methods 0.000 claims description 2
- 230000000694 effects Effects 0.000 abstract description 3
- 238000013519 translation Methods 0.000 description 6
- 238000000354 decomposition reaction Methods 0.000 description 3
- 230000000052 comparative effect Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 241001028048 Nicola Species 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000012827 research and development Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
- G06F16/367—Ontology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/35—Clustering; Classification
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Animal Behavior & Ethology (AREA)
- Computational Linguistics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention overcomes the problem that the existing time-aware knowledge map embedding method cannot learn the duration distribution rule of the similar knowledge, and provides a time-aware knowledge representation learning method integrating hyperplane and duration modeling. The method comprises the steps of classifying each relation event in a knowledge graph, modeling the duration of continuous knowledge, calculating the effective reliability of knowledge, using the effective reliability as training weight, improving the knowledge graph embedding expression learning process, and obtaining an embedding model of the knowledge graph. Through effective reliability calculation and weighting training in embedded representation learning, the obtained model can learn the duration distribution rule of the similar knowledge, and therefore the accuracy of knowledge map entities and the relation link prediction task is improved. The method can be used for improving the embedding and representing effect of the knowledge graph of time perception and improving the accuracy of knowledge graph link prediction.
Description
Technical Field
The invention relates to a knowledge graph, in particular to an embedded representation method of the knowledge graph, which is mainly suitable for the link prediction of the knowledge graph of time perception.
Background
The knowledge map stores human knowledge in a structured mannerThe knowledge system is a knowledge base with a directed graph structure, and is a universal formal description framework of semantic knowledge. Knowledge representation learning
[1]The method is the basis of knowledge acquisition and application and is an important problem throughout the whole process of knowledge map construction and application. Although the traditional structural triple representation form has strong expression capability, the semantic calculation is difficult to be directly carried out by using a computer. The embedding expression method expresses discrete symbols (entities, attributes, relations, values and the like) in the knowledge graph by continuous numerical values, directly reflects semantic information, can efficiently calculate the entities, the relations and complex semantic associations thereof, and greatly improves the efficiency of semantic calculation of the knowledge graph. The embedded representation method of the current knowledge can be roughly divided into: tensor-based, translation-based, etc.
Tensor-based approach: the method utilizes tensor to express knowledge and uses tensor decomposition or tensor multiplication to express and learn knowledge, and has the advantages of integrating the information of the whole knowledge map in the process of coding entities and relations and having strong expression capability. Due to the fact that the calculation amount of tensor decomposition reconstruction is large, the knowledge representation learning efficiency of the large knowledge graph based on the tensor decomposition method is low.
Translation-based methods: the method takes the relation in the triple as a translation vector from a head entity to a tail entity, the training aims to enable the translation of the head entity vector through the relation vector to be close to the vector of the tail entity, the problem of low learning efficiency is solved, the representation learning can be rapidly completed in a large-scale knowledge graph, and the most classical is TransE
[2]A method. The invention relates to a knowledge representation method based on translation.
The traditional translation-based method assumes that the knowledge graph exists under the same time-space condition, and ignores an important implied semantic information of the knowledge graph, namely time information. Therefore, in recent years, the use of time information for knowledge representation learning has rapidly become a research focus in the field of knowledge representation learning. The current time-aware knowledge representation learning method mainly comprises time sequence-based knowledge representation, time coding-based knowledge representation, time hyperplane-based knowledge representation and the likeHyperplanar knowledge representation method (HyTE)
[3]) The method is a fast and effective method, and has the core idea that the entity and the relation are mapped onto a time hyperplane through time information, an evaluation function is calculated, and the time information is creatively and directly embedded into a hyperplane space, so that the link prediction accuracy of the entity and the relation is effectively improved, and wide attention is drawn. However, the knowledge representation method based on the time hyperplane has the following problems:
(1) the effective duration of knowledge is cut into independent time slices by the knowledge representation method based on the time hyperplane, so that the time slices are split, the effective duration lengths of the same kind of knowledge are similar, and a model trained by the method based on the time hyperplane cannot effectively learn the knowledge of the time lengths.
(2) In the knowledge graph containing the time labels, the confidence level of knowledge at each time point is different, the more distant the time point is from the starting time point, the lower the confidence level of the knowledge is effective, and the information cannot be effectively applied in the training process of the time hyperplane-based method, so that the link prediction precision of the model is influenced.
Aiming at the problems, the invention uses a duration modeling method for reference to model the duration of each relation event in the knowledge graph, calculates the effective credibility of knowledge at each time point, improves a knowledge representation learning method based on a time hyperplane by taking the effective credibility as a training weight, and provides a knowledge graph embedded representation learning method combining the hyperplane and the effective credibility so as to improve the accuracy of knowledge graph link prediction of time perception.
Disclosure of Invention
The invention aims to improve the accuracy of prediction of the time-aware knowledge graph entity link and the relationship link. Effective credibility of knowledge is calculated through effective duration modeling, and the effective credibility is used as a training process of a weight optimization model, the training weight of high-credibility knowledge is improved, so that the purposes of optimizing a prediction model and improving the model link prediction accuracy are achieved, and the main contents comprise:
i. efficient confidence computation process
1) And (5) knowledge classification. All knowledge is first classified according to its duration, mainly into persistent and transient relationships. A persistent relationship is a relationship that persists for a certain period of time, such relationship being characterized by a non-zero duration; transient-type relationships refer to relationships that occur at a certain time and that do not persist.
2) And calculating effective credibility. The duration distribution of the persistence type relationship satisfies a certain rule, and is generally close to the gaussian distribution, so that the duration distribution can be rapidly simulated by using the gaussian distribution. And modeling duration of each continuous knowledge by taking Gaussian distribution as a reliable function, and calculating the effective credibility of each knowledge on different time slices.
A training process of knowledge-graph embedded representation models.
First, vectors are randomly generated for each entity and relationship, and each time slice is represented by a hyperplane. When an evaluation function is calculated, the entity and the relation vector are projected onto a time hyperplane, and the effective credibility of the knowledge on the time slice is used as a weight to optimize a model training process. And continuously adjusting the entity and the relation vector through a gradient descent method, and optimizing the embedded representation model.
Entity to relationship linking prediction process.
And learning semantic information in the knowledge graph as much as possible by using the model obtained by training the learning algorithm. And testing the entity lacking the entity or the relation in the embedded representation model, calculating an evaluation function, and obtaining the predicted entity and relation according to the score of the evaluation function.
Drawings
FIG. 1 is a general framework diagram of the algorithm proposed by the present invention;
FIG. 2 is a graph of duration distribution of persistent knowledge;
FIG. 3 is a graph of instantaneous knowledge duration distribution;
FIG. 4 is a graph comparing the Mean Rank curves predicted for the tail entity;
FIG. 5 is a graph comparing the predicted Hits @10 (%) variation curves for the tail entity;
FIG. 6 is a graph comparing the relationship-predicted Mean Rank curves;
FIG. 7 is a graph comparing the curves for the dependence of the predicted Hits @1 (%);
Detailed Description
The invention is described in detail below with reference to the figures and the examples.
The invention adopts public data sets YAG011K and WikiData12K as test data. The method comprises the steps of firstly calculating effective credibility of knowledge on different time slices based on duration modeling, then improving a training process of knowledge graph embedding representation learning by taking the effective credibility as a weight to obtain an embedding representation model, and finally performing entity and relationship prediction on the knowledge lacking head and tail entities and relationships to obtain predicted entity and relationship links. The general block diagram of the invention is shown in fig. 1, and the implementation process is as follows:
the method comprises the following steps:
step 1: dividing all knowledge in the knowledge map into instantaneous knowledge and continuous knowledge according to duration;
step 2: carrying out duration modeling on the duration knowledge, and calculating the effective reliability VR of each effective time slice;
step 3: initializing all entities, relations and time vectors randomly;
step 4: calculating an evaluation function by taking VR as weight, and adjusting each entity, relation and time vector;
step 5: repeating step4 until reaching preset cycle times, and outputting final entity, relation and time vector;
step 6: and predicting missing entities and relations by taking the output entities, relations and time vectors as prediction models, and taking the average ranking (Mean Rank) of correct entities and relations and the percentage of correct results in the top ten (Hits @ 10%) and the top one (Hits @ 1%) as evaluation indexes.
The method comprises the following specific steps:
step 1: knowledge classification
One of the knowledge-graphs contains a time-tagged meta-fact quadruple (h, r, t, [ tau ] of
s，τ
e]) Recording a meta fact (h, r, t) in the time dimensionThe process from invalid to valid to invalid (invalid → valid → invalid), where (h, r, t, [ tau ]
s，τ
e]) Representation (head subject entity, predicate relationship, tail object entity, [ start time, end time [ ]]) Is a record of the knowledge graph. The division criteria for persistent versus transient relationships are:
fig. 2 and fig. 3 show the duration distribution of two kinds of continuous knowledge and instantaneous knowledge, respectively, and it can be found that: persistent knowledge refers to knowledge that is continuously effective for a certain period of time, and is characterized by non-zero duration; instantaneous knowledge refers to knowledge that occurs at a certain time and is not persistent.
Step 2: knowledge validation confidence computation
The effective credibility refers to a quadruple (h, r, t, [ tau ] in a certain knowledge graph containing time labels
s，τ
e]) From the start time (tau)
s) After t time, the probability that the knowledge is still Valid is called Valid Reliability (VR) and is denoted as VR. And according to the duration distribution of the duration relation, performing duration modeling on each type of duration knowledge, and calculating the effective credibility of the knowledge. In order to facilitate the modeling of the duration of a large amount of knowledge in a large knowledge graph, a duration modeling method of a Gaussian function capable of rapidly simulating effective duration distribution is adopted. Thus, the probability density function f of knowledge can be expressed as:
wherein sigma
rRefers to the standard deviation of the duration of all knowledge containing the relationship r. For a quadruple (h, r, t, [ tau ]
s，，τ
e]) Duration of τ
pThe cumulative distribution function F is expressed as
Wherein sigma
rMeaning the standard deviation, τ, of the duration of the meta-facts containing the relation r
eRefers to an end time point of the valid time of the event. From the above two equations, the knowledge duration of the persistence relationship can be derived as τ
pEffective confidence of the time:
unlike the persistent relationship, the duration of the transient relationship is 0, so the effective confidence level of knowledge containing the transient relationship is 1 at the effective time and 0 at the rest of the time. In addition, since the negative example is incorrect knowledge in the knowledge-graph, whenever it is incorrect validity is true. The effective confidence of a negative sample as incorrect knowledge is therefore 1 at any point in time. In combination with the above discussion of the positive and negative sample valid confidence value problem, the positive and negative sample valid confidence calculation formula can be expressed as:
and step 3: embedded representation model training process
The method is a knowledge representation learning method of time perception, and when an evaluation function is calculated, the evaluation function is firstly mapped onto a time plane. Time is represented by a hyperplane. For the T time slices, we denote by the normal vectors of T different time hyperplanes. The merit function may be expressed as:
wherein P is
τ(e
？) Representing the vectors resulting from projecting the head, tail or relationship vectors onto the temporal hyperplane at time point τ. The embedded expression of the knowledge graph requires negative sampling to accelerate the training process and improve the model link prediction precisionThe method adopts a negative sampling method which does not count time:
wherein D
+And ξ represents an entity set for the positive sample set, and the evaluation function calculation of the positive and negative samples is integrated, and the calculation formula of the model training loss function is as follows:
and 4, step 4: entity and relationship linkage prediction process
For each missing head and tail entity or relationship knowledge, firstly projecting it onto a corresponding time hyperplane, the formula of the entity and relationship projection on the time hyperplane is:
wherein P is
τ(e
？) Representing the vectors resulting from projecting the head, tail or relationship vectors onto the temporal hyperplane at time point τ. h, r, t respectively represent head entity (head entry), relationship (relationship) and tail entity (tail entry), e
？Representing its corresponding vector.
The results of the evaluation function of all candidate entities or relationships are then calculated, and the scores for all entities or relationships are ranked. The evaluation function is calculated by the formula:
through the time hyperplane projection and the evaluation function calculation, the evaluation index of the prediction effect can be further calculated. There are three evaluation indexes of the link prediction effect: the percentage of entities with correct results in the top ten names is called Hits @ 10%, the percentage of relations with correct results in the first name is called Hits @ 1%, and the average ranking of correct results is Mean Rank.
Fig. 4 and 5 are comparative experimental results of HyTE on the wikitata 12K data set and the tail entity link prediction task in the present invention and the conventional hyperplane-based method, respectively, and fig. 6 and 7 are comparative experimental results of HyTE on the wikitata 12K data set and the related prediction task in the present invention and the conventional hyperplane-based method, respectively. Therefore, after knowledge classification, knowledge effective reliability calculation and embedded representation model training, the method obviously improves the entity link prediction and relationship link prediction precision of the knowledge graph.
Reference to the literature
[1] Liu zhi, sun louping, lin yan kai, xie luo ice knowledge represents the progress of learning research [ J ] computer research and development, 2016, 53 (02): 247-261.
[2]Antoine Bordes，Nicolas Usunier，Alberto GarciaDuran，Jason Weston，and Oksana Yakhnenko.2013.Translating embeddings for modeling multirelationaldata.In C.J.C.Burges，L. Bottou，M.Welling，Z.Ghahramani，and K.Q.Weinberger，editors，Advances in Neural Information Processing Systems 26，pages 2787-2795.Curran Associates，Inc.
[3]Shib Sankar Dasgupta，Swayambhu Nath Ray，Partha Talukdar.(2018).HyTE：Hyperplane- based Temporally aware Knowledge Graph Embedding.2001-2011.D18-1225。
Claims (6)
1. The time perception knowledge representation learning method integrating hyperplane and duration modeling is characterized by comprising the following steps of:
1) the embedding expression method based on the knowledge graph is characterized in that time is expressed by a hyperplane normal vector;
2) defining an instantaneous relationship and a continuous relationship according to duration;
3) defining the effective credibility of the knowledge, and modeling and calculating the effective credibility of the knowledge by the duration;
4) the time-aware knowledge representation learning algorithm is optimized.
2. The method as claimed in claim 1, wherein the time-aware knowledge representation learning method fusing hyperplane and duration modeling is characterized in that the knowledge graph-based embedding representation method is used, time is represented in a hyperplane form, before the model is trained, an instantaneous relationship and a duration relationship are defined according to the duration of knowledge, all knowledge is classified according to the duration of the relationship, then the effective credibility of the knowledge is defined, the duration modeling is carried out on the knowledge of the duration relationship, the effective credibility of the knowledge is calculated, and finally the effective credibility of the knowledge is used as a training weight for optimizing the training algorithm of the knowledge graph embedding representation model. Compared with other knowledge graph embedding expression learning methods, the algorithm can optimize the model training process and improve the entity link prediction precision, the relation prediction precision and the time prediction precision of the model.
3. The method of time-aware knowledge representation learning fusing hyperplane and duration modeling as claimed in claim 1, wherein both duration-type knowledge and transient-type knowledge are defined. One of the knowledge-graphs contains a time-tagged meta-fact quadruple (h, r, t, [ tau ] of
s，τ
e]) Recording the process of a meta fact (h, r, t) from invalid to valid to invalid (invalid → valid → invalid) in the time dimension, wherein (h, r, t, [ tau ]
s，τ
e]) Representation (head subject entity, predicate relationship, tail object entity, [ start time, end time [ ]]) Is a record of the knowledge graph. The division criteria for persistent versus transient relationships are:
persistent knowledge refers to knowledge that is continuously effective for a certain period of time, and is characterized by non-zero duration; instantaneous knowledge refers to knowledge that occurs at a certain time and is not persistent.
4. The method of time-aware knowledge representation learning fusing hyperplane and duration modeling as claimed in claim 1, wherein an effective confidence level of knowledge is defined. The effective credibility refers to a quadruple (h, r, t, [ tau ] in a certain knowledge graph containing time labels
s，τ
e]) From the start time (tau)
s) After t time, the probability that the knowledge is still Valid is called Valid Reliability (VR) and is denoted as VR. And according to the duration distribution of the duration relation, performing duration modeling on each type of duration knowledge, and calculating the effective credibility of the knowledge. In order to facilitate the modeling of the duration of a large amount of knowledge in a large knowledge graph, a duration modeling method of a Gaussian function capable of rapidly simulating effective duration distribution is adopted. Thus, the probability density function f of knowledge can be expressed as:
where σ r refers to the standard deviation of the duration of all knowledge containing the relationship r. For a quadruple (h, r, t, [ tau ]
s，，τ
e]) Duration of τ
pThe cumulative distribution function F is expressed as
Where σ r refers to the standard deviation of the duration of the meta-facts containing the relation r, τ
eRefers to an end time point of the valid time of the event. From the above two equations, the knowledge duration of the persistence relationship can be derived as τ
pEffective confidence of the time:
unlike the persistent relationship, the duration of the transient relationship is 0, so the effective confidence level of knowledge containing the transient relationship is 1 at the effective time and 0 at the rest of the time. In addition, since the negative example is incorrect knowledge in the knowledge-graph, whenever it is incorrect validity is true. The effective confidence of a negative sample as incorrect knowledge is therefore 1 at any point in time. In combination with the above discussion of the positive and negative sample valid confidence value problem, the positive and negative sample valid confidence calculation formula can be expressed as:
。
5. the method of time-aware knowledge representation learning fusing hyperplane and duration modeling as claimed in claim 1, wherein the process of model training is improved. The method is a knowledge representation learning method of time perception, and when an evaluation function is calculated, the evaluation function is firstly mapped onto a time plane. Time is represented by a hyperplane. For the T time slices, we denote by the normal vectors of T different time hyperplanes. The merit function may be expressed as:
wherein P is
τ(e
？) Representing the vectors resulting from projecting the head, tail or relationship vectors onto the temporal hyperplane at time point τ. The embedding of the knowledge graph shows that negative sampling is needed to accelerate the training process and improve the model link prediction precision, and the negative sampling method adopted by the method is an untimely negative sampling method:
wherein D
+For the positive sample set, ξ represents the entity setCalculating an evaluation function of the positive and negative samples, wherein a calculation formula of a model training loss function is as follows:
6. the method for learning a time-aware knowledge representation incorporating hyperplane and duration modeling as claimed in claim 1, wherein the algorithm trained on the model is optimized. Firstly, modeling and calculating effective credibility through knowledge effective duration, initializing a model randomly, then calculating a loss function by taking the effective credibility as a training weight, and continuously updating the model, wherein the method comprises the following specific steps:
step 1: dividing all knowledge in the knowledge map into instantaneous knowledge and continuous knowledge according to duration;
step 2: carrying out duration modeling on the duration knowledge, and calculating the effective reliability VR of each effective time slice;
step 3: initializing all entities, relations and time vectors randomly;
step 4: calculating an evaluation function by taking VR as weight, and adjusting each entity, relation and time vector;
step 5: repeating step4 until reaching the preset cycle number, and outputting the final entity, relation and time vector;
step 6: and predicting missing entities and relations by taking the output entities, relations and time vectors as prediction models, and taking the average ranking (Mean Rank) of correct entities and relations and the percentage of correct results in the top ten (Hits @ 10%) and the top one (Hits @ 1%) as evaluation indexes.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911020394.6A CN110781316A (en) | 2019-10-25 | 2019-10-25 | Time perception knowledge representation learning method integrating hyperplane and duration modeling |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911020394.6A CN110781316A (en) | 2019-10-25 | 2019-10-25 | Time perception knowledge representation learning method integrating hyperplane and duration modeling |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110781316A true CN110781316A (en) | 2020-02-11 |
Family
ID=69387760
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911020394.6A Pending CN110781316A (en) | 2019-10-25 | 2019-10-25 | Time perception knowledge representation learning method integrating hyperplane and duration modeling |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110781316A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113032580A (en) * | 2021-03-29 | 2021-06-25 | 浙江星汉信息技术股份有限公司 | Associated file recommendation method and system and electronic equipment |
| CN113742491A (en) * | 2021-08-12 | 2021-12-03 | 上海熙业信息科技有限公司 | Representation learning-based time knowledge graph reasoning method |
| CN113946695A (en) * | 2021-12-20 | 2022-01-18 | 山东新希望六和集团有限公司 | Method and device for generating animal pedigree and computer equipment |
| CN114117064A (en) * | 2021-11-09 | 2022-03-01 | 西南交通大学 | Knowledge dynamic evolution method based on multi-time granularity and application |
| CN115292523A (en) * | 2022-08-04 | 2022-11-04 | 中国科学院空天信息创新研究院 | Spatiotemporal information reasoning method based on graph representation learning |
-
2019
- 2019-10-25 CN CN201911020394.6A patent/CN110781316A/en active Pending
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113032580A (en) * | 2021-03-29 | 2021-06-25 | 浙江星汉信息技术股份有限公司 | Associated file recommendation method and system and electronic equipment |
| CN113742491A (en) * | 2021-08-12 | 2021-12-03 | 上海熙业信息科技有限公司 | Representation learning-based time knowledge graph reasoning method |
| CN114117064A (en) * | 2021-11-09 | 2022-03-01 | 西南交通大学 | Knowledge dynamic evolution method based on multi-time granularity and application |
| CN113946695A (en) * | 2021-12-20 | 2022-01-18 | 山东新希望六和集团有限公司 | Method and device for generating animal pedigree and computer equipment |
| CN115292523A (en) * | 2022-08-04 | 2022-11-04 | 中国科学院空天信息创新研究院 | Spatiotemporal information reasoning method based on graph representation learning |
| CN115292523B (en) * | 2022-08-04 | 2023-09-22 | 中国科学院空天信息创新研究院 | Spatial-temporal information reasoning method based on graph representation learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110781316A (en) | Time perception knowledge representation learning method integrating hyperplane and duration modeling | |
| CN110928764B (en) | Automated evaluation method for crowdsourcing test report of mobile application and computer storage medium | |
| CN112699377A (en) | Function-level code vulnerability detection method based on slice attribute graph representation learning | |
| CN106557563B (en) | Query statement recommendation method and device based on artificial intelligence | |
| CN113590645B (en) | Searching method, searching device, electronic equipment and storage medium | |
| CN113535984A (en) | Attention mechanism-based knowledge graph relation prediction method and device | |
| CN110334186B (en) | Data query method and device, computer equipment and computer readable storage medium | |
| EP3493109A1 (en) | System and methods for faster processor comparisons of visual graph features | |
| CN113128206B (en) | Question generation method based on word importance weighting | |
| CN113313229A (en) | Bayesian optimization of sparsity in model compression | |
| CN112949433B (en) | Method, device and equipment for generating video classification model and storage medium | |
| CN114722833A (en) | Semantic classification method and device | |
| CN111967187B (en) | Modeling method and equipment for predicting output data of starting process of gas turbine | |
| CN114298299A (en) | Model training method, device, equipment and storage medium based on course learning | |
| CN113642727A (en) | Training method of neural network model and processing method and device of multimedia information | |
| CN114757097B (en) | Line fault diagnosis method and device | |
| CN116307792B (en) | Urban physical examination subject scene-oriented evaluation method and device | |
| CN110852066B (en) | Multi-language entity relation extraction method and system based on confrontation training mechanism | |
| CN116306923A (en) | Evaluation weight calculation method based on knowledge graph | |
| CN110399382A (en) | Civil aviaton's master data recognition methods and system based on cloud model and rough set | |
| CN113420680A (en) | Remote sensing image area attention and text generation method based on GRU attention | |
| Wu et al. | Application in Computer Software Testing Based on Artificial Intelligence Technology | |
| CN116484053B (en) | Intelligent data analysis platform | |
| Dubovski et al. | Left barrier loss for unbiased survival analysis prediction | |
| CN114842246A (en) | Social media pressure category detection method and device |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20200211 |
|
| WD01 | Invention patent application deemed withdrawn after publication |