CN108647258A - A kind of expression learning method based on entity associated constraint - Google Patents
A kind of expression learning method based on entity associated constraint Download PDFInfo
- Publication number
- CN108647258A CN108647258A CN201810377516.6A CN201810377516A CN108647258A CN 108647258 A CN108647258 A CN 108647258A CN 201810377516 A CN201810377516 A CN 201810377516A CN 108647258 A CN108647258 A CN 108647258A
- Authority
- CN
- China
- Prior art keywords
- entity
- batch
- strong
- formula
- positive sample
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Machine Translation (AREA)
Abstract
The present invention relates to a kind of expression learning methods based on entity associated constraint, belong to natural language processing and knowledge mapping technical field.The present invention is explained by the description text to entity and relevance divides, obtain the strong associated entity set and weak rigidity entity sets of entity, it is fused to relevance as auxiliary loss item in the expression learning method based on translation, in particular by the negative sampling of sample and model training, obtain the embedded expression of entity and relationship, i.e. by knowledge mapping head entity h, tail entity t and head entity and tail entity between relationship r, be respectively embedded in vectorial h, vector t and vector r.The method of the invention is in reasoning effect better than the expression learning method based on translation and based on text model.
Description
Technical field
The present invention relates to a kind of expression learning methods based on entity associated constraint, belong to natural language processing and know
Know graphical spectrum technology field.
Background technology
Knowledge mapping (Knowledge Graph) is a kind of knowledge representation method based on semantic network, provides one kind
Efficient and succinct structured representation mode plays key effect in Web search and intelligent answer field.Knowledge mapping will be true
The data in the real world are expressed as entity and relationship, and knowledge passes through relationship with (entity, relationship, entity) triple store between entity
It is coupled to constitute the netted structure of knowledge.Although knowledge mapping contains a large amount of entity and relationship, due to knowledge mapping
All facts can not be drawn into building process, it is complete to lead to knowledge mapping not.Knowledge mapping inference technology can answer
It can also be taken out with the information of Opening field such as relationship that may be present between two entities of prediction for the automation completion of collection of illustrative plates
It takes and is combined, quality evaluation is carried out to information extraction result.Indicate learning method refer to by knowledge mapping entity and relationship it is embedding
Enter to a lower dimensional space, knowledge mapping reasoning is completed in lower dimensional space.
Relationship is regarded as a kind of translation process between entity, the referred to as mould based on translation by the expression learning method of mainstream at present
Type, such as TransE models (Bordes A, Usunier N, Weston J, et al.Translating embeddings for
modeling multi-relational data[C]in International Conference on Neural
Information Processing Systems.Curran Associates Inc.2013:2787-2795) propose relationship
It is the translating operation between an entity and tail entity, which only makes inferences from the structure of figure, can not provide more auxiliary
The room for promotion of supplementary information, reasoning effect is limited.Then scholar proposes the table for combining entity description text and graph structure
Dendrography learning method belongs to text representation model, therein to represent DKRL models (Xie, R., Liu, Z., Jia, J., Luan, H., ＆
Sun,M.(2016,February).Representation Learning of Knowledge Graphs with Entity
Descriptions.In AAAI.pp.2659-2665) by description text and structural information the progress combination learning of entity, it uses
Bag of words (CBOW) and convolutional neural networks (CNN) encode text, however but having ignored can between entity in text
Semantic association existing for energy can lose more semantic information in training.
Although existing expression learning method achieves good effect in knowledge mapping reasoning, to real in text
Semantic association between body, which lacks, to be excavated, and there is also much rooms for the improvement in reasoning performance.The purpose of the present invention is be dedicated to solving
Certainly tradition indicates that learning method can lose the technological deficiency of the semantic association information between entity in training, it is proposed that one kind is based on
The expression learning method of entity associated constraint.
Invention content
It is an object of the invention to not utilize semantic information abundant in text, Yi Jiwen for the model based on translation
This expression model can lose the semantic association information between entity in the training process, it is proposed that one kind is constrained based on entity associated
Expression learning method.
Core of the invention thought is：Go out relevance entity based on entity description text mining and relevance is divided
Grade, using relevance as in auxiliary Constraint fusion to the expression learning method based on translation；The main entity using after explaining is retouched
It states text and obtains the co-occurrence information between entity, standard of the information as semantic association degree between a kind of two entities of measurement, and
And the correlation degree is oriented.Specific implementation is by head entity h, tail entity t and the head entity and tail reality in knowledge mapping
Relationship r between body is respectively embedded in vectorial h, vector t and vector r.
The present invention is realized by following steps：
Step 1: carrying out note and relevance division to the description text of entity, the strong associated entity set of entity is obtained
With weak rigidity entity sets；It include specifically following sub-step：
Step 1.1, the description text annotations to entity, obtain entity annotation result；
Wherein, entity refers to the entity in knowledge mapping, is indicated with e；The description text of e, uses DeseIt indicates, being one section has
The set of letters of sequence is indicated with formula (1)：
Dese=＜ w1,...,wm＞ (1)
Wherein, w1,...,wmIt is word, m is the quantity for describing word in text, and the entity for describing to extract in text is by big
It is formed in equal to 1 word, when entity is formed by being more than or equal to two words, needs to spell the word extracted
It connects；
The process of extraction entity is known as describing text annotations from description text；The entity extracted in text composition will be described
Set is to get to entity annotation result：
Dese'=＜ w1,...,wm'＞ (2)
Wherein, m'≤m, wiIndicate an entity, Dese' it is DeseEntity annotation result；
Step 1.2, relevance divide；
J pairs of entity is obtained by formula (3) using i-th in the entity annotation result of step 1.1 output and j-th of entity
The correlation degree value of entity i, uses WijIt indicates：
If Wij=2, the strong associated entity (Strong Relevant Entity) that note j is i；If Wij=1, note j are
The weak rigidity entity (Weak Relevant Entity) of i, if two entities mutually occur in describing each other, relevance
Become strong, then obtains the strong associated entity set and weak rigidity entity sets of entity e；
Correlation degree value WijIt is oriented, traverses all entities in entity annotation result, obtains correlation degree value composition
Entity associated matrix, be denoted asE is the entity sets in knowledge mapping, | E | indicate the entity in knowledge mapping
Total number：
Wherein, the strong associated entity set of entity e is denoted as S (e)：
Wherein, eiI-th of entity is represented,Represent entity e and entity eiBetween be strong associated entity relationship；
The weak rigidity entity sets of entity e are denoted as W (e)：
Step 2: the negative sampling of sample and model training, obtain the embedded expression of entity and relationship；It represents real
Body e and entity eiBetween be weak rigidity entity relationship；
Wherein, model training is the training carried out based on batch stochastic gradient descent algorithm；
Step 2 specifically includes following sub-step：
Step 2.1, loop initialization count value are 1 and cycle count maximum value；
Wherein, loop count is denoted as k；Cycle count maximum value, is denoted as iter；
Step 2.2 enables S indicate the triplet sets in knowledge mapping, a triple in knowledge mapping be one just
Sample, i.e. S are positive sample set；B positive sample is randomly selected from S obtains a subset conjunction Sbatch, enable
TbatchConstruction include following sub-step：
Step 2.2.1, S is traversedbatch, negative sampling is carried out to each positive sample (h, r, t), bears the method for sampling with document 1
(Feng J.Knowledge Graph Embedding by Translating on Hyperplanes[C]in
AAAI.2014)：Given relationship r, sample is born average per tail entity corresponding to head entity described in the method for sampling in document 1
Number tph corresponds to the tph of this patentr, the head number of entities hpt corresponding to average each tail entity corresponds to the hpt of this patentr；
The equally distributed random number p in [0, a 1] section is generated, if p is less than or equal to tphr/(tphr+hptr), then
An entity is extracted from the entity sets E equal probabilities of knowledge mapping and replaces head entity in positive sample, and is ensured after replacing
Triple be not belonging to S；If p is more than tphr/(tphr+hptr), then it is taken out from the entity sets E equal probabilities of knowledge mapping
It takes an entity to replace the tail entity in positive sample, and ensures that replaced triple is not belonging to S；
Step 2.2.2, after the completion of replacement, S can be obtainedbatchIn negative sample corresponding to each positive sample (h, r, t)
Each positive sample and negative sample are added to T by (h', r, t')batchIn set：
Tbatch←Tbatch∪{(h,r,t),(h',r,t')} (6)
T is obtained after step 2.2.1 and step 2.2.2batchSet extracts T outbatchIn entity sets, be denoted as
Ebatch；
Step 2.3 is trained model based on batch stochastic gradient descent algorithm；
Step 2.3.1, by formula (7), the score function of triple (h, r, t) calculates the scoring of triple (h, r, t),
It is denoted as fr(h,t)；
Wherein,Represent square of 2 norms of h+r-t vectors；
Step 2.3.2, the loss item L based on entity associated is calculated by formula (8)r：
Wherein, α and β is strong associated weight value and weak rigidity weights, and α determines the intensity of High relevancy constraint, and β determines weak
The intensity of relevance constraint；E represents EbatchIn entity；On the left side of formula (8), e' represents the strong associated entity set of e,
On the right of formula (8), the weak rigidity entity sets of e' tables e；Square of 2 norms of representation vector e-e'；SC
The strong distance limit for being associated with hyper parameter and weak rigidity hyper parameter, indicating between two associated entities respectively that user specifies is represented with WC
System, when entity is in corresponding range, loss is 0, LrSo that related entity does not surpass the distance in vector space
A certain range is crossed, and not associated entity is made to adjust the distance minimum simply；
Step 2.3.3, pass through the loss function value of formula (9) computation model：
Wherein, the loss function value of Loss representative models；fr(h, t) represents the scoring of positive sample (h, r, t), fr(h',t')
The scoring of negative sample (h', r, t') is represented, positive sample scoring can be made to tend to low value when training, negative sample scoring tends to high level；
γ is that loss is spaced, and γ is for controlling fr(h, t) and frThe difference of (h', t')；
Step 2.3.4, the derivative in calculation formula (9) about independent variable, and be updated according to formula (10)；
Wherein, θ is independent variable, including all h, r and t, and rate is learning rate,Represent the loss letter to model
Numerical value Loss takes differential about independent variable θ；
Step 2.3.5, judge whether loop count k has reached count maximum iter, if k=iter, complete
This method；Otherwise k=k+1 skips to step 2.2；
So far, from step 1 to step 2, the embedded expression of entity and relationship has been obtained：Vectorial h, vector t and vector
R completes a kind of expression learning method constrained based on entity associated.
Advantageous effect
It is a kind of based on entity associated constraint expression learning method have the advantages that compared with the conventional method：
1. the structured message of the more knowledge based collection of illustrative plates of traditional expression learning method, does not make full use of the description of entity
Text message, the present invention propose a kind of associated method of measurement Entity Semantics from entity description text, use this method structure
A kind of bound term based on entity associated is made, which, which is fused to tradition, indicates in learning method, experimental result table
Bright, the present invention compares conventional method reasoning effect more in the link prediction task and ternary component generic task on public data collection
It is good and consistent with conventional method in speed；
2. text based indicates that learning method mostly by text vector, has ignored the semantic association between entity in text,
Entity associated bound term proposed by the present invention is excavated from the entity in text, more meticulously the semantic association between entity
It is modeled, experimental result indicates that the present invention compares text representation model DKRL, and reasoning effect is more preferable.
Description of the drawings
Fig. 1 is the flow signal in a kind of expression learning method constrained based on entity associated of the present invention and embodiment 1
Figure.
Specific implementation mode
The present invention will be further described with reference to the accompanying drawings and examples and detailed description.
Embodiment 1
The present embodiment describes a kind of specific implementation of the expression learning method constrained based on entity associated of the present invention
Journey, the implementation process schematic diagram of Fig. 1 the present embodiment.
From figure 1 it appears that steps are as follows for the specific implementation of the present invention and the present embodiment:
Step A, to the description text of entity carry out explain and relevance divide, obtain entity strong associated entity set and
Weak rigidity entity sets；It include specifically following sub-step：
Step A.1, to the description text annotations of entity, obtain entity annotation result；
Wherein, entity refers to the entity in knowledge mapping, is indicated with e；The description text of e, uses DeseIt indicates, being one section has
The set of letters of sequence is indicated with formula (1)：
Dese=＜ w1,...,wm＞ (11)
Wherein, w1,...,wmIt is word, m is the quantity for describing word in text, and the entity for describing to extract in text is by big
It is formed in equal to 1 word, when entity is formed by being more than or equal to two words, needs to spell the word extracted
It connects；
The process of extraction entity is known as describing text annotations from description text；The entity extracted in text composition will be described
Set is to get to entity annotation result：
Dese'=＜ w1,...,wm'＞ (12)
Wherein, m'≤m, wiIndicate an entity, Dese' it is DeseEntity annotation result；
Step A.2, relevance divide；
I-th and j-th of entity obtain j pairs of entity by formula (3) in the entity annotation result A.1 exported using step
The correlation degree value of entity i, uses WijIt indicates：
If Wij=2, the strong associated entity (Strong Relevant Entity) that note j is i；If Wij=1, note j are
The weak rigidity entity (Weak Relevant Entity) of i, if two entities mutually occur in describing each other, relevance
Become strong, then obtains the strong associated entity set and weak rigidity entity sets of entity e；
Correlation degree value WijIt is oriented, traverses all entities in entity annotation result, obtain entity associated matrix, remembers
ForE is the entity sets in knowledge mapping, | E | indicate the entity total number in knowledge mapping：
Wherein, the strong associated entity set of entity e is denoted as S (e)：
The weak rigidity entity sets of entity e are denoted as W (e)：
Step B, the negative sampling of sample and model training, obtain the embedded expression of entity and relationship；
Wherein, model training is the training carried out based on batch stochastic gradient descent algorithm；Step 2 specifically includes as follows
Sub-step：
Step B.1, loop initialization count value, wherein loop count is denoted as k, initializes k=1；
Step B.2, enable S indicate the triplet sets in knowledge mapping, a triple in knowledge mapping be one just
Sample, i.e. S are positive sample set；B positive sample is randomly selected from S obtains a subset conjunction Sbatch, wherein B takes 100, enablesTbatchConstruction include following sub-step：
Step B.2.1, traversal Sbatch, negative sampling is carried out to each positive sample (h, r, t), bears the method for sampling with document 1
(Feng J.Knowledge Graph Embedding by Translating on Hyperplanes[C]in
AAAI.2014)：Given relationship r, sample is born average per tail entity corresponding to head entity described in the method for sampling in document 1
Number tph corresponds to the tph of this patentr, the head number of entities hpt corresponding to average each tail entity corresponds to the hpt of this patentr；
The equally distributed random number p in [0, a 1] section is generated, if p is less than or equal to tphr/(tphr+hptr), then
An entity is extracted from the entity sets E equal probabilities of knowledge mapping and replaces head entity in positive sample, and is ensured after replacing
Triple be not belonging to S；If p is more than tphr/(tphr+hptr), then it is taken out from the entity sets E equal probabilities of knowledge mapping
It takes an entity to replace the tail entity in positive sample, and ensures that replaced triple is not belonging to S；
Step B.2.2, replace after the completion of, S can be obtainedbatchIn negative sample corresponding to each positive sample (h, r, t)
Each positive sample and negative sample are added to T by (h', r, t')batchIn set：
Tbatch←Tbatch∪{(h,r,t),(h',r,t')} (16)
By step B.2.1 with step B.2.2 after obtain TbatchSet extracts T outbatchIn entity sets, be denoted as
Ebatch；
B.3, based on batch stochastic gradient descent algorithm step is trained model；
Step B.3.1, by formula (7), the score function of triple (h, r, t) calculates the scoring of triple (h, r, t),
It is denoted as fr(h,t)；
Wherein,Represent square of 2 norms of h+r-t vectors；
B.3.2, by formula (8) step calculates the loss item L based on entity associatedr：
Wherein, α and β is strong associated weight value and weak rigidity weights, and SC and WC are respectively strong association range and weak rigidity range,
Wherein α=1, β=0.3, SC=1, WC=1；
Step 2.3.3, pass through the loss function value of formula (9) computation model：
Wherein, γ is that loss is spaced, and γ takes 1；
Step 2.3.4, the derivative in calculation formula (9) about independent variable, and be updated according to formula (10)；
Wherein, θ is independent variable, including all h, r and t, and rate is learning rate, rate=0.1；
Step 2.3.5, judge whether loop count k has reached count maximum iter, iter=500, if k=
Iter completes this method；Otherwise k=k+1 skips to step 2.1.
So far, from step 1 to step 2, the embedded expression of entity and relationship has been obtained：Vectorial h, vector t and vector
R completes a kind of expression learning method constrained based on entity associated.
Claims (2)
1. a kind of expression learning method based on entity associated constraint, it is characterised in that：Core concept is：Based on entity description
Text mining goes out relevance entity and is classified to relevance, using relevance as auxiliary Constraint fusion to the table based on translation
In dendrography learning method；The main entity description text using after explaining obtains the co-occurrence information between entity, and the information is as a kind of
The standard of semantic association degree between two entities of measurement, and the correlation degree is oriented；Specific implementation is by knowledge mapping
In head entity h, tail entity t and head entity and tail entity between relationship r, be respectively embedded in vectorial h, vector t and vector r
In；It is realized especially by following steps：
Step 1: to the description text of entity explain and relevance divides, the strong associated entity set of entity and weak is obtained
Associated entity set；It include specifically following sub-step：
Step 1.1, the description text annotations to entity, obtain entity annotation result；
Wherein, entity refers to the entity in knowledge mapping, is indicated with e；The description text of e, uses DeseIt indicates, is one section orderly
Set of letters is indicated with formula (1)：
Dese=＜ w1,...,wm＞ (1)
Wherein, w1,...,wmIt is word, m is the quantity for describing word in text, and the entity for describing to extract in text is by being more than
It is formed in 1 word, when entity is formed by being more than or equal to two words, needs to splice the word extracted；
The process of extraction entity is known as describing text annotations from description text；It will describe the entity extracted in text and form collection
It closes to get to entity annotation result：
Dese'=＜ w1,...,wm'＞ (2)
Wherein, m'≤m, wiIndicate an entity, Dese' it is DeseEntity annotation result；
Step 1.2, relevance divide；
Entity j is obtained to entity by formula (3) using i-th in the entity annotation result of step 1.1 output and j-th of entity
The correlation degree value of i, uses WijIt indicates：
If Wij=2, the strong associated entity (Strong Relevant Entity) that note j is i；If Wij=1, note j is the weak of i
Associated entity (Weak Relevant Entity), if two entities mutually occur in describing each other, relevance becomes strong,
Then the strong associated entity set and weak rigidity entity sets of entity e are obtained；
All entities in entity annotation result are traversed, the entity associated matrix of correlation degree value composition is obtained, is denoted asE is the entity sets in knowledge mapping, | E | indicate the entity total number in knowledge mapping：
Wherein, the strong associated entity set of entity e is denoted as S (e)：
Wherein, eiI-th of entity is represented,Represent entity e and entity eiBetween be strong associated entity relationship；
The weak rigidity entity sets of entity e are denoted as W (e)：
Step 2: the negative sampling of sample and model training, obtain the embedded expression of entity and relationship；Represent entity e and
Entity eiBetween be weak rigidity entity relationship, specifically include following sub-step：
Step 2.1, loop initialization count value are 1 and cycle count maximum value；
Wherein, loop count is denoted as k；Cycle count maximum value, is denoted as iter；
Step 2.2 enables S indicate the triplet sets in knowledge mapping, and a triple in knowledge mapping is a positive sample,
That is S is positive sample set；B positive sample is randomly selected from S obtains a subset conjunction Sbatch, enableTbatchStructure
It makes including following sub-step：
Step 2.2.1, S is traversedbatch, negative sampling is carried out to each positive sample (h, r, t), bears the method for sampling with 1 (Feng of document
J.Knowledge Graph Embedding by Translating on Hyperplanes[C]in AAAI.2014)：It gives
Determine relationship r, sample bears the average number tph correspondences per tail entity corresponding to head entity described in the method for sampling originally in document 1
The tph of patentr, the head number of entities hpt corresponding to average each tail entity corresponds to the hpt of this patentr；
The equally distributed random number p in [0, a 1] section is generated, if p is less than or equal to tphr/(tphr+hptr), then from knowing
It extracts the head entity in an entity replacement positive sample with knowing the entity sets E equal probabilities of collection of illustrative plates, and ensures replaced three
Tuple is not belonging to S；If p is more than tphr/(tphr+hptr), then extract one from the entity sets E equal probabilities of knowledge mapping
A entity replaces the tail entity in positive sample, and ensures that replaced triple is not belonging to S；
Step 2.2.2, after the completion of replacement, S can be obtainedbatchIn corresponding to each positive sample (h, r, t) negative sample (h', r,
T'), each positive sample and negative sample are added to TbatchIn set：
Tbatch←Tbatch∪{(h,r,t),(h',r,t')} (6)
T is obtained after step 2.2.1 and step 2.2.2batchSet extracts T outbatchIn entity sets, be denoted as Ebatch；
Step 2.3 is trained model based on batch stochastic gradient descent algorithm；
Step 2.3.1, by formula (7), the score function of triple (h, r, t) calculates the scoring of triple (h, r, t), is denoted as
fr(h,t)；
Wherein,Represent square of 2 norms of h+r-t vectors；
Step 2.3.2, the loss item L based on entity associated is calculated by formula (8)r：
Wherein, α and β is strong associated weight value and weak rigidity weights, and α determines that the intensity of High relevancy constraint, β determine weak rigidity
Property constraint intensity；E represents EbatchIn entity；On the left side of formula (8), e' represents the strong associated entity set of e, in public affairs
The right of formula (8), the weak rigidity entity sets of e' tables e；Square of 2 norms of representation vector e-e'；SC and WC
The strong association hyper parameter and weak rigidity hyper parameter that user specifies are represented, indicates the distance limitation between two associated entities respectively,
When entity is in corresponding range, loss is 0, LrSo that related entity is no more than the distance in vector space
A certain range, and not associated entity is made to adjust the distance minimum simply；
Step 2.3.3, pass through the loss function value of formula (9) computation model：
Wherein, the loss function value of Loss representative models；fr(h, t) represents the scoring of positive sample (h, r, t), fr(h', t') is represented
The scoring of negative sample (h', r, t'), can make positive sample scoring tend to low value when training, negative sample scoring tends to high level；γ is
Loss interval, γ is for controlling fr(h, t) and frThe difference of (h', t')；
Step 2.3.4, the derivative in calculation formula (9) about independent variable, and be updated according to formula (10)；
Wherein, θ is independent variable, including all h, r and t, and rate is learning rate,Represent the loss function value to model
Loss takes differential about independent variable θ；
Step 2.3.5, judge whether loop count k has reached count maximum iter, if k=iter, complete we
Method；Otherwise k=k+1 skips to step 2.2；
So far, from step 1 to step 2, the embedded expression of entity and relationship has been obtained：Vectorial h, vector t and vector r.
2. a kind of expression learning method based on entity associated constraint according to claim 1, it is characterised in that：Association
Degree value WijIt is oriented.
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2018100675556 | 2018-01-24 | ||
| CN201810067555 | 2018-01-24 |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108647258A true CN108647258A (en) | 2018-10-12 |
| CN108647258B CN108647258B (en) | 2020-12-22 |
Family
ID=63747612
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810377516.6A Active CN108647258B (en) | 2018-01-24 | 2018-04-25 | Representation learning method based on entity relevance constraint |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108647258B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110674637A (en) * | 2019-09-06 | 2020-01-10 | 腾讯科技（深圳）有限公司 | Character relation recognition model training method, device, equipment and medium |
| CN110909881A (en) * | 2019-11-01 | 2020-03-24 | 中电科大数据研究院有限公司 | Knowledge representation method for cross-media knowledge reasoning task |
| CN111428047A (en) * | 2020-03-19 | 2020-07-17 | 东南大学 | Knowledge graph construction method and device based on UC L semantic indexing |
| CN113220833A (en) * | 2021-05-07 | 2021-08-06 | 支付宝(杭州)信息技术有限公司 | Entity association degree identification method and device |
| CN114330323A (en) * | 2022-03-08 | 2022-04-12 | 成都数联云算科技有限公司 | Entity relationship joint extraction method and device, computer terminal and storage medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105630901A (en) * | 2015-12-21 | 2016-06-01 | 清华大学 | Knowledge graph representation learning method |
| US20160154876A1 (en) * | 2010-06-07 | 2016-06-02 | Microsoft Technology Licensing, Llc | Using context to extract entities from a document collection |
| CN107122399A (en) * | 2017-03-16 | 2017-09-01 | 中国科学院自动化研究所 | Combined recommendation system based on Public Culture knowledge mapping platform |
| CN107273349A (en) * | 2017-05-09 | 2017-10-20 | 清华大学 | A kind of entity relation extraction method and server based on multilingual |
-
2018
- 2018-04-25 CN CN201810377516.6A patent/CN108647258B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160154876A1 (en) * | 2010-06-07 | 2016-06-02 | Microsoft Technology Licensing, Llc | Using context to extract entities from a document collection |
| CN105630901A (en) * | 2015-12-21 | 2016-06-01 | 清华大学 | Knowledge graph representation learning method |
| CN107122399A (en) * | 2017-03-16 | 2017-09-01 | 中国科学院自动化研究所 | Combined recommendation system based on Public Culture knowledge mapping platform |
| CN107273349A (en) * | 2017-05-09 | 2017-10-20 | 清华大学 | A kind of entity relation extraction method and server based on multilingual |
Non-Patent Citations (3)
| Title |
|---|
| HAN XIAO等: "SSP：semantic space projection for knowledge graph embedding with text descriptions", 《THIRTY-FIRST AAAI CONFERENCE ON ARTIFICIAL INTELLIGENCE》 * |
| RUOBING XIE等: "Representation learning of knowledge graphs with entity descriptions", 《THIRTIETH AAAI CONFERENCE ON ARTIFICIAL INTELLIGENCE》 * |
| 刘峤等: "知识图谱构建技术综述", 《计算机研究与发展》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110674637A (en) * | 2019-09-06 | 2020-01-10 | 腾讯科技（深圳）有限公司 | Character relation recognition model training method, device, equipment and medium |
| CN110674637B (en) * | 2019-09-06 | 2023-07-11 | 腾讯科技（深圳）有限公司 | Character relationship recognition model training method, device, equipment and medium |
| CN110909881A (en) * | 2019-11-01 | 2020-03-24 | 中电科大数据研究院有限公司 | Knowledge representation method for cross-media knowledge reasoning task |
| CN110909881B (en) * | 2019-11-01 | 2022-11-04 | 中电科大数据研究院有限公司 | Knowledge representation method for cross-media knowledge reasoning task |
| CN111428047A (en) * | 2020-03-19 | 2020-07-17 | 东南大学 | Knowledge graph construction method and device based on UC L semantic indexing |
| CN111428047B (en) * | 2020-03-19 | 2023-04-21 | 东南大学 | Knowledge graph construction method and device based on UCL semantic indexing |
| CN113220833A (en) * | 2021-05-07 | 2021-08-06 | 支付宝(杭州)信息技术有限公司 | Entity association degree identification method and device |
| CN114330323A (en) * | 2022-03-08 | 2022-04-12 | 成都数联云算科技有限公司 | Entity relationship joint extraction method and device, computer terminal and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108647258B (en) | 2020-12-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108647258A (en) | A kind of expression learning method based on entity associated constraint | |
| CN110825881B (en) | Method for establishing electric power knowledge graph | |
| CN109558487A (en) | Document Classification Method based on the more attention networks of hierarchy | |
| CN109635291A (en) | A kind of recommended method of fusion score information and item contents based on coorinated training | |
| CN108388651A (en) | A kind of file classification method based on the kernel of graph and convolutional neural networks | |
| CN107526799A (en) | A kind of knowledge mapping construction method based on deep learning | |
| CN110245229A (en) | A kind of deep learning theme sensibility classification method based on data enhancing | |
| CN110232186A (en) | The knowledge mapping for merging entity description, stratification type and text relation information indicates learning method | |
| CN108509425A (en) | Chinese new word discovery method based on novelty | |
| CN107766371A (en) | A kind of text message sorting technique and its device | |
| CN109033129A (en) | Multi-source Information Fusion knowledge mapping based on adaptive weighting indicates learning method | |
| CN108549658A (en) | A kind of deep learning video answering method and system based on the upper attention mechanism of syntactic analysis tree | |
| CN108509654A (en) | The construction method of dynamic knowledge collection of illustrative plates | |
| CN110415071B (en) | Automobile competitive product comparison method based on viewpoint mining analysis | |
| CN109165275B (en) | Intelligent substation operation ticket information intelligent search matching method based on deep learning | |
| CN110162631A (en) | Chinese patent classification method, system and storage medium towards TRIZ inventive principle | |
| CN108399241A (en) | A kind of emerging much-talked-about topic detecting system based on multiclass feature fusion | |
| CN107145514A (en) | Chinese sentence pattern sorting technique based on decision tree and SVM mixed models | |
| CN113360582B (en) | Relation classification method and system based on BERT model fusion multi-entity information | |
| CN109299248A (en) | A kind of business intelligence collection method based on natural language processing | |
| CN109858008A (en) | The tendentious method and device of document court verdict based on deep learning | |
| CN114741519A (en) | Paper correlation analysis method based on graph convolution neural network and knowledge base | |
| CN108664652A (en) | Expression learning method based on Path selection under a kind of complex network | |
| CN113886562A (en) | AI resume screening method, system, equipment and storage medium | |
| CN110851733A (en) | Community discovery and emotion interpretation method based on network topology and document content |
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 |