CN108647258B - Representation learning method based on entity relevance constraint - Google Patents

Representation learning method based on entity relevance constraint Download PDF

Info

Publication number
CN108647258B
CN108647258B CN201810377516.6A CN201810377516A CN108647258B CN 108647258 B CN108647258 B CN 108647258B CN 201810377516 A CN201810377516 A CN 201810377516A CN 108647258 B CN108647258 B CN 108647258B
Authority
CN
China
Prior art keywords
entity
entities
relevance
batch
vector
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.)
Active
Application number
CN201810377516.6A
Other languages
Chinese (zh)
Other versions
CN108647258A (en
Inventor
刘琼昕
马敬
龙航
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Institute of Technology BIT
Original Assignee
Beijing Institute of Technology BIT
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Beijing Institute of Technology BIT filed Critical Beijing Institute of Technology BIT
Publication of CN108647258A publication Critical patent/CN108647258A/en
Application granted granted Critical
Publication of CN108647258B publication Critical patent/CN108647258B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

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 invention relates to an expression learning method based on entity relevance constraint, and belongs to the technical field of natural language processing and knowledge maps. According to the method, a strong association entity set and a weak association entity set of an entity are obtained by performing annotation and association division on a description text of the entity, the association is used as an auxiliary loss item and is fused into a translation-based representation learning method, and specifically, embedded representation of the entity and the relationship is obtained through sample negative sampling and model training, namely a head entity h and a tail entity t in a knowledge graph and a relationship r between the head entity and the tail entity are respectively embedded into a vector h, the vector t and the vector r. The method is superior to representation learning methods based on translation and text models in reasoning effect.

Description

Representation learning method based on entity relevance constraint
Technical Field
The invention relates to an expression learning method based on entity relevance constraint, and belongs to the technical field of natural language processing and knowledge maps.
Background
A Knowledge Graph (Knowledge Graph) is a Knowledge representation method based on a semantic network, provides an efficient and concise structural representation mode, and plays a key role in the fields of Web search and intelligent question answering. The knowledge graph represents real world data as entities and relations, knowledge is stored in (entities, relations, entities) triples, and the entities are connected through relations to form a network knowledge structure. Although a knowledge graph contains a large number of entities and relationships, the knowledge graph is not complete due to the fact that all of the knowledge graph cannot be extracted during the construction process of the knowledge graph. The knowledge graph reasoning technology can be applied to automatic completion of the graph, for example, the possible relation between two entities is predicted, and the quality evaluation can be carried out on the information extraction result by combining with the information extraction in the open field. The expression learning method is to embed the entities and relations in the knowledge graph into a low-dimensional space and finish the inference of the knowledge graph in the low-dimensional space.
Currently, the mainstream expression learning method regards the relationship as a translation process between entities, which is called a translation-based model, such as a trans model (border a, user N, Weston J, et al. translation expressions for modeling multi-translation data [ C ] in International Conference on Neural Information Processing systems, current associations inc.2013: 2787) which proposes that the relationship is a translation operation between a head entity and a tail entity. Therefore, some researchers propose a Representation Learning method combining an Entity description text and a graph structure, which belongs to a text Representation model, wherein a Representation DKRL model (Xie, r., Liu, z., Jia, j., Luan, h., & Sun, M. (2016, library). reproduction Learning of Knowledge Graphs with Entity descriptions.
Although the existing representation learning method has good effect on knowledge graph reasoning, semantic association between entities in a text is lack of mining, and a great space exists for improving reasoning performance. The invention aims to overcome the technical defect that semantic association information among entities is lost in the training process of the traditional expression learning method, and provides an expression learning method based on entity association constraint.
Disclosure of Invention
The invention aims to provide a representation learning method based on entity relevance constraint aiming at the problems that a translation-based model does not utilize rich semantic information in a text and the text representation model loses semantic relevance information among entities in the training process.
The core idea of the invention is as follows: mining out a relevance entity based on the entity description text, grading the relevance, and fusing the relevance serving as an auxiliary constraint into a translation-based representation learning method; the annotated entity description text is mainly used for obtaining the co-occurrence information between the entities, and the information is used as a standard for measuring the semantic association degree between the two entities, and the association degree is directional. The specific implementation is that a head entity h, a tail entity t and a relation r between the head entity and the tail entity in the knowledge graph are respectively embedded into a vector h, a vector t and a vector r.
The invention is realized by the following steps:
firstly, annotating and performing relevance division on a description text of an entity to obtain a strong relevance entity set and a weak relevance entity set of the entity; the method specifically comprises the following substeps:
step 1.1, annotating description texts of entities to obtain entity annotation results;
wherein the entity refers to an entity in the knowledge graph and is represented by e; e description text by DeseThe expression is an ordered word set, and is expressed by formula (1):
Dese＝＜w1,...,wm＞ (1)
wherein, w1,...,wmThe words are used, m is the number of words in the description text, the entity extracted from the description text is composed of more than or equal to 1 word, and when the entity is composed of more than or equal to two words, the extracted words need to be spliced;
the process of extracting entities from descriptive text is called descriptive text annotation; and (3) forming a set by the entities extracted from the description text to obtain an entity annotation result:
Dese'＝＜w1,...,wm'＞ (2)
wherein m' is less than or equal to m, wiRepresents an entity, DeseIs' DeseThe entity annotation result of (1);
step 1.2, relevance division;
obtaining the association degree value of the entity j to the entity i by using the ith entity and the jth entity in the entity annotation result output in the step 1.1 through a formula (3), and using WijRepresents:
Figure BDA0001640177200000031
if W isij2, a strongly associated Entity (Strong Relevant Entity) denoted j as i; if W isij1, recording j as a Weak associated Entity (Weak relevance Entity) of i, if two entities appear in the description of each other, the association becomes strong, and then obtaining a strong associated Entity set and a Weak associated Entity set of an Entity e;
degree of association value WijTraversing all entities in the entity annotation result to obtain an entity incidence matrix composed of incidence degree values, and recording the entity incidence matrix as
Figure BDA0001640177200000032
E is the set of entities in the knowledge-graph, | E | represents the total number of entities in the knowledge-graph:
wherein, the strongly associated entity set of the entity e is denoted as s (e):
Figure BDA0001640177200000033
wherein e isiWhich represents the (i) th entity, the (ii) th entity,
Figure BDA0001640177200000034
representing entity e and entity eiThe relationship between the two is strong association entity;
the weakly associated entity set of entity e is denoted as w (e):
Figure BDA0001640177200000041
secondly, carrying out sample negative sampling and model training to obtain an embedded expression of an entity and a relation;
Figure BDA0001640177200000042
representing entity e and entity eiThe relationship between the two is weak association entity;
the model training is based on batch random gradient descent algorithm;
step two, the following substeps are specifically included:
step 2.1, initializing a cycle count value to be 1 and a cycle count maximum value;
wherein, the cycle count value is marked as k; the maximum cycle count, denoted iter;
step 2.2, enabling S to represent a triple set in the knowledge graph, wherein one triple in the knowledge graph is a positive sample, namely S is a positive sample set; randomly extracting B positive samples from S to obtain a subset SbatchLet us order
Figure BDA0001640177200000043
TbatchThe construction of (1) comprises the following substeps:
step 2.2.1, traverse SbatchEach positive sample (h, r, t) was sampled negatively as described in document 1(Feng J. knowledge Graph Embedding by transformation on Hyperplanes [ C ]]in aaai.2014): given the relation r, the average number tph of tail entities corresponding to each head entity in the sample negative sampling method in document 1 corresponds to tph of the patentrThe average number hpt of head entities corresponding to each tail entity corresponds to hpt of the patentr；
Produce a [0,1 ]]Uniformly distributed random number p of interval, if p is less than or equal to tphr/(tphr+hptr) Then equal probability is given from the set of entities E of the knowledge-graphAn entity is extracted to replace a head entity in the positive sample, and the replaced triple is ensured not to belong to S; if p is greater than tphr/(tphr+hptr) Extracting an entity from the entity set E of the knowledge graph with equal probability to replace the tail entity in the positive sample, and ensuring that the replaced triple does not belong to S;
step 2.2.2, after the replacement is completed, S can be obtainedbatchA negative sample (h ', r, T') corresponding to each positive sample (h, r, T), adding each positive sample and negative sample to TbatchIn the set:
Tbatch←Tbatch∪{(h,r,t),(h',r,t')} (6)
after the steps 2.2.1 and 2.2.2, T is obtainedbatchSet, extract TbatchThe entity set in (1), denoted as Ebatch；
2.3, training the model based on a batch random gradient descent algorithm;
step 2.3.1, calculating the scores of the triplets (h, r, t) through the scoring function of the triplets (h, r, t) in the formula (7), and marking the scores as fr(h,t)；
Figure BDA0001640177200000051
Wherein,
Figure BDA0001640177200000055
represents the square of the 2 norm of the h + r-t vector;
step 2.3.2, calculating a loss term L based on entity relevance through a formula (8)r：
Figure BDA0001640177200000052
Wherein, alpha and beta are a strong correlation weight and a weak correlation weight, alpha determines the strength of the strong correlation constraint, and beta determines the strength of the weak correlation constraint; e represents EbatchThe entity of (1); on the left hand side of equation (8), e' represents the strongly associated set of entities for e, in equation(8) The right entry of (e), the weakly associated entity set of table e;
Figure BDA0001640177200000053
represents the square of the 2 norm of the vector e-e'; SC and WC represent strong correlation hyper-parameters and weak correlation hyper-parameters specified by a user, respectively represent distance limits between two associated entities, and when an entity pair is in a corresponding range, the loss is 0, and L isrSuch that the distance of the associated entity pair in vector space does not exceed a certain range and does not monotonically minimize the associated entity pair distance;
and 2.3.3, calculating a loss function value of the model according to a formula (9):
Figure BDA0001640177200000054
wherein, Loss represents the Loss function value of the model; f. ofr(h, t) represents the score of a positive sample (h, r, t), fr(h ', t') represents the scores of the negative samples (h ', r, t'), and the positive samples tend to be low in score and the negative samples tend to be high in score when training; gamma is the loss interval, gamma is used to control fr(h, t) and fr(h ', t');
step 2.3.4, calculating the derivative of the independent variable in the formula (9) and updating according to the formula (10);
Figure BDA0001640177200000061
wherein theta is an independent variable, all h, r and t are included, rate is a learning rate,
Figure BDA0001640177200000062
representing the differentiation of the Loss function value Loss of the model with respect to the argument θ;
step 2.3.5, judging whether the cycle count value k reaches the maximum value iter, and if k is iter, finishing the method; otherwise k is k +1, jumping to step 2.2;
so far, from step one to step two, an embedded representation of entities and relationships is obtained: and completing a representation learning method based on entity relevance constraint by using the vector h, the vector t and the vector r.
Advantageous effects
Compared with the prior art, the representation learning method based on entity relevance constraint has the following beneficial effects:
1. the traditional representation learning method is based on the structural information of the knowledge graph, and the description text information of the entity is not fully utilized, the invention provides a method for measuring the semantic association of the entity from the entity description text, a constraint item based on the entity association is constructed by using the method, the constraint item is fused into the traditional representation learning method, and the experimental result shows that the invention has better reasoning effect on the link prediction task and the triple classification task on the open data set compared with the traditional method, and the speed is consistent with that of the traditional method;
2. the expression learning method based on the text mostly vectorizes the text, omits semantic association among entities in the text, excavates entity association constraint items from the entities in the text, more carefully models the semantic association among the entities, and the experimental result shows that the expression learning method based on the text has better reasoning effect compared with a text expression model DKRL.
Drawings
Fig. 1 is a schematic flow diagram of a representation learning method based on entity association constraint and an embodiment 1 of the invention.
Detailed Description
The invention is further illustrated and described in detail below with reference to the figures and examples.
Example 1
This embodiment describes a specific implementation process of the representation learning method based on entity association constraint according to the present invention, and fig. 1 is an implementation flow diagram of this embodiment.
As can be seen from fig. 1, the specific implementation steps of the present invention and the embodiment are as follows:
a, annotating and performing relevance division on a description text of an entity to obtain a strong relevance entity set and a weak relevance entity set of the entity; the method specifically comprises the following substeps:
a.1, annotating description texts of entities to obtain entity annotation results;
wherein the entity refers to an entity in the knowledge graph and is represented by e; e description text by DeseThe expression is an ordered word set, and is expressed by formula (1):
Dese＝＜w1,...,wm＞ (11)
wherein, w1,...,wmThe words are used, m is the number of words in the description text, the entity extracted from the description text is composed of more than or equal to 1 word, and when the entity is composed of more than or equal to two words, the extracted words need to be spliced;
the process of extracting entities from descriptive text is called descriptive text annotation; and (3) forming a set by the entities extracted from the description text to obtain an entity annotation result:
Dese'＝＜w1,...,wm'＞ (12)
wherein m' is less than or equal to m, wiRepresents an entity, DeseIs' DeseThe entity annotation result of (1);
step A.2, relevance division;
obtaining the association degree value of the entity j to the entity i by the ith entity and the jth entity in the entity annotation result output in the step A.1 through a formula (3), and using WijRepresents:
Figure BDA0001640177200000071
if W isij2, a strongly associated Entity (Strong Relevant Entity) denoted j as i; if W isij1, recording j as a Weak associated Entity (Weak relevance Entity) of i, if two entities appear in the description of each other, the association becomes strong, and then obtaining a strong associated Entity set and a Weak associated Entity set of an Entity e;
degree of association value WijIs directed, traverses entity annotationsAll entities in the result obtain an entity incidence matrix, which is marked as
Figure BDA0001640177200000081
E is the set of entities in the knowledge-graph, | E | represents the total number of entities in the knowledge-graph:
wherein, the strongly associated entity set of the entity e is denoted as s (e):
Figure BDA0001640177200000082
the weakly associated entity set of entity e is denoted as w (e):
Figure BDA0001640177200000083
b, carrying out sample negative sampling and model training to obtain an embedded expression of an entity and a relation;
the model training is based on batch random gradient descent algorithm; step two, the following substeps are specifically included:
step B.1, initializing a cycle count value, wherein the cycle count value is recorded as k, and the initialization k is 1;
b.2, enabling S to represent a triple set in the knowledge graph, wherein one triple in the knowledge graph is a positive sample, namely S is a positive sample set; randomly extracting B positive samples from S to obtain a subset SbatchWherein B is 100, 1
Figure BDA0001640177200000084
TbatchThe construction of (1) comprises the following substeps:
step B.2.1, traverse SbatchEach positive sample (h, r, t) was sampled negatively as described in document 1(Feng J. knowledge Graph Embedding by transformation on Hyperplanes [ C ]]in aaai.2014): given the relation r, the average number tph of tail entities corresponding to each head entity in the sample negative sampling method in document 1 corresponds to tph of the patentrAverage of each pair of tail entitiesThe number hpt of head entities corresponds to hpt of the patentr；
Produce a [0,1 ]]Uniformly distributed random number p of interval, if p is less than or equal to tphr/(tphr+hptr) Extracting an entity from the entity set E of the knowledge graph with equal probability to replace the head entity in the positive sample, and ensuring that the replaced triple does not belong to S; if p is greater than tphr/(tphr+hptr) Extracting an entity from the entity set E of the knowledge graph with equal probability to replace the tail entity in the positive sample, and ensuring that the replaced triple does not belong to S;
step B.2.2, after the replacement is finished, S can be obtainedbatchA negative sample (h ', r, T') corresponding to each positive sample (h, r, T), adding each positive sample and negative sample to TbatchIn the set:
Tbatch←Tbatch∪{(h,r,t),(h',r,t')} (16)
obtaining T after the step B.2.1 and the step B.2.2batchSet, extract TbatchThe entity set in (1), denoted as Ebatch；
B.3, training the model based on a batch random gradient descent algorithm;
and B.3.1, calculating the scores of the triples (h, r, t) through the scoring function of the triples (h, r, t) in the formula (7), and marking the scores as fr(h,t)；
Figure BDA0001640177200000091
Wherein,
Figure BDA0001640177200000092
represents the square of the 2 norm of the h + r-t vector;
step B.3.2, calculating a loss term L based on entity relevance through a formula (8)r：
Figure BDA0001640177200000093
Wherein α and β are a strong association weight and a weak association weight, SC and WC are a strong association range and a weak association range, respectively, where α ═ 1, β ═ 0.3, SC ═ 1, and WC ═ 1;
and 2.3.3, calculating a loss function value of the model according to a formula (9):
Figure BDA0001640177200000094
wherein gamma is a loss interval and is 1;
step 2.3.4, calculating the derivative of the independent variable in the formula (9) and updating according to the formula (10);
Figure BDA0001640177200000095
wherein, θ is an independent variable including all h, r and t, rate is a learning rate, and rate is 0.1;
step 2.3.5, judging whether the cycle count value k reaches the maximum value iter, wherein iter is 500, and if k is iter, completing the method; otherwise k is k +1, go to step 2.1.
So far, from step one to step two, an embedded representation of entities and relationships is obtained: and completing a representation learning method based on entity relevance constraint by using the vector h, the vector t and the vector r.

Claims (1)

1. A representation learning method based on entity relevance constraint is characterized in that: the core idea is as follows: mining out a relevance entity based on the entity description text, grading the relevance, and fusing the relevance serving as an auxiliary constraint into a translation-based representation learning method; obtaining co-occurrence information between entities by mainly utilizing the annotated entity description text, wherein the information is used as a standard for measuring semantic association degree between two entities, and the association degree is directional; the specific implementation is that a head entity h, a tail entity t and a relation r between the head entity and the tail entity in a knowledge graph are respectively embedded into a vector h, a vector t and a vector r; the method is realized by the following steps:
firstly, annotating and performing relevance division on a description text of an entity to obtain a strong relevance entity set and a weak relevance entity set of the entity; the method specifically comprises the following substeps:
step 1.1, annotating description texts of entities to obtain entity annotation results;
wherein the entity refers to an entity in the knowledge graph and is represented by e; e description text by DeseThe expression is an ordered word set, and is expressed by formula (1):
Dese＝＜w1,...,wm＞ (1)
wherein, w1,...,wmThe words are used, m is the number of words in the description text, the entity extracted from the description text is composed of more than or equal to 1 word, and when the entity is composed of more than or equal to two words, the extracted words need to be spliced;
the process of extracting entities from descriptive text is called descriptive text annotation; and (3) forming a set by the entities extracted from the description text to obtain an entity annotation result:
Dese'＝＜w1,...,wm'＞ (2)
wherein m' is less than or equal to m, wiRepresents an entity, DeseIs' DeseThe entity annotation result of (1);
step 1.2, relevance division;
obtaining the association degree value of the entity j to the entity i by using the ith entity and the jth entity in the entity annotation result output in the step 1.1 through a formula (3), and using WijRepresents:
Figure FDA0002719609540000021
if W isij2, a strongly associated Entity (Strong Relevant Entity) denoted j as i; if W isijA weakly associated Entity (Weak relevance) denoted j by i, which becomes strongly associated if two entities appear to each other in the description of each otherThen, a strong associated entity set and a weak associated entity set of the entity e are obtained;
traversing all entities in the entity annotation result to obtain an entity association matrix consisting of association degree values, and marking as W e ∈|E|×|E|E is the set of entities in the knowledge-graph, | E | represents the total number of entities in the knowledge-graph:
wherein, the strongly associated entity set of the entity e is denoted as s (e):
Figure FDA0002719609540000022
wherein e isiWhich represents the (i) th entity, the (ii) th entity,
Figure FDA0002719609540000023
representing entity e and entity eiThe relationship between the two is strong association entity;
the weakly associated entity set of entity e is denoted as w (e):
Figure FDA0002719609540000024
secondly, carrying out sample negative sampling and model training to obtain an embedded expression of an entity and a relation;
Figure FDA0002719609540000025
representing entity e and entity eiThe method specifically comprises the following substeps:
step 2.1, initializing a cycle count value to be 1 and a cycle count maximum value;
wherein, the cycle count value is marked as k; the maximum cycle count, denoted iter;
step 2.2, enabling S to represent a triple set in the knowledge graph, wherein one triple in the knowledge graph is a positive sample, namely S is a positive sample set; randomly extracting B positive samples from S to obtain a subset SbatchLet us order
Figure FDA0002719609540000026
The construction of (1) comprises the following substeps:
step 2.2.1, traverse Sbatch-negative sampling each positive sample (h, r, t);
produce a [0,1 ]]Uniformly distributed random number p of interval, if p is less than or equal to tphr/(tphr+hptr) Extracting an entity from the entity set E of the knowledge graph with equal probability to replace the head entity in the positive sample, and ensuring that the replaced triple does not belong to S; if p is greater than tphr/(tphr+hptr) Extracting an entity from the entity set E of the knowledge graph with equal probability to replace the tail entity in the positive sample, and ensuring that the replaced triple does not belong to S;
step 2.2.2, after the replacement is completed, S can be obtainedbatchA negative sample (h ', r, T') corresponding to each positive sample (h, r, T), adding each positive sample and negative sample to TbatchIn the set:
Figure FDA0002719609540000031
after the steps 2.2.1 and 2.2.2, T is obtainedbatchSet, extract TbatchThe entity set in (1), denoted as Ebatch；
2.3, training the model based on a batch random gradient descent algorithm;
step 2.3.1, calculating the scores of the triplets (h, r, t) through the scoring function of the triplets (h, r, t) in the formula (7), and marking the scores as fr(h,t)；
Figure FDA0002719609540000032
Wherein,
Figure FDA0002719609540000033
representing 2 norms of a h + r-t vectorSquaring;
step 2.3.2, calculating a loss term L based on entity relevance through a formula (8)r：
Figure FDA0002719609540000034
Wherein, alpha and beta are a strong correlation weight and a weak correlation weight, alpha determines the strength of the strong correlation constraint, and beta determines the strength of the weak correlation constraint; e represents EbatchThe entity of (1); in the left term of formula (8), e 'represents the strongly associated entity set of e, and in the right term of formula (8), e' represents the weakly associated entity set of e;
Figure FDA0002719609540000035
represents the square of the 2 norm of the vector e-e'; SC and WC represent strong correlation hyper-parameters and weak correlation hyper-parameters specified by a user, respectively represent distance limits between two associated entities, and when an entity pair is in a corresponding range, the loss is 0, and L isrSuch that the distance of the associated entity pair in vector space does not exceed a certain range and does not monotonically minimize the associated entity pair distance;
and 2.3.3, calculating a loss function value of the model according to a formula (9):
Figure FDA0002719609540000041
wherein, Loss represents the Loss function value of the model; f. ofr(h, t) represents the score of a positive sample (h, r, t), fr(h ', t') represents the scores of the negative samples (h ', r, t'), and the positive samples tend to be low in score and the negative samples tend to be high in score when training; gamma is the loss interval, gamma is used to control fr(h, t) and fr(h ', t');
step 2.3.4, calculating the derivative of the independent variable in the formula (9) and updating according to the formula (10);
Figure FDA0002719609540000042
wherein theta is an independent variable, all h, r and t are included, rate is a learning rate,
Figure FDA0002719609540000043
representing the differentiation of the Loss function value Loss of the model with respect to the argument θ;
step 2.3.5, judging whether the cycle count value k reaches the maximum value iter, and if k is iter, finishing the method; otherwise k is k +1, jumping to step 2.2;
so far, from step one to step two, an embedded representation of entities and relationships is obtained: vector h, vector t, and vector r.
CN201810377516.6A 2018-01-24 2018-04-25 Representation learning method based on entity relevance constraint Active CN108647258B (en)

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 CN108647258A (en) 2018-10-12
CN108647258B true 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)

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
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
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
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
CN114330323B (en) * 2022-03-08 2022-06-28 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 (3)

* Cited by examiner, † Cited by third party
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
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

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9251248B2 (en) * 2010-06-07 2016-02-02 Microsoft Licensing Technology, LLC Using context to extract entities from a document collection

Patent Citations (3)

* Cited by examiner, † Cited by third party
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
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)

* Cited by examiner, † Cited by third party
Title
Representation learning of knowledge graphs with entity descriptions;Ruobing Xie等;《Thirtieth AAAI Conference on Artificial Intelligence》;20161231;2659-2665 *
SSP：semantic space projection for knowledge graph embedding with text descriptions;Han Xiao等;《Thirty-First AAAI conference on artificial intelligence》;20171231;3014-3110 *
知识图谱构建技术综述;刘峤等;《计算机研究与发展》;20161231;第53卷(第03期);582-600 *

Also Published As

Publication number Publication date
CN108647258A (en) 2018-10-12

Similar Documents

Publication Publication Date Title
CN108647258B (en) Representation learning method based on entity relevance constraint
CN113254599B (en) Multi-label microblog text classification method based on semi-supervised learning
CN108536870B (en) Text emotion classification method fusing emotional features and semantic features
CN110569508A (en) Method and system for classifying emotional tendencies by fusing part-of-speech and self-attention mechanism
CN105975573B (en) A kind of file classification method based on KNN
CN111144448A (en) Video barrage emotion analysis method based on multi-scale attention convolutional coding network
CN109635109A (en) Sentence classification method based on LSTM and combination part of speech and more attention mechanism
CN112395393B (en) Remote supervision relation extraction method based on multitask and multiple examples
CN110619121B (en) Entity relation extraction method based on improved depth residual error network and attention mechanism
CN110298403A (en) The sentiment analysis method and system of enterprise dominant in a kind of financial and economic news
CN111581385A (en) Chinese text type identification system and method for unbalanced data sampling
CN111274814B (en) Novel semi-supervised text entity information extraction method
CN112395417A (en) Network public opinion evolution simulation method and system based on deep learning
CN111506728B (en) Hierarchical structure text automatic classification method based on HD-MSCNN
CN110334187A (en) Burmese sentiment analysis method and device based on transfer learning
CN111914555B (en) Automatic relation extraction system based on Transformer structure
CN113723083A (en) Weighted negative supervision text emotion analysis method based on BERT model
CN113886562A (en) AI resume screening method, system, equipment and storage medium
Chu et al. Co-training based on semi-supervised ensemble classification approach for multi-label data stream
CN115630156A (en) Mongolian emotion analysis method and system fusing Prompt and SRU
CN114548117A (en) Cause-and-effect relation extraction method based on BERT semantic enhancement
CN113011192B (en) Text emotion feature extraction method based on attention causal interpretation
CN116050419B (en) Unsupervised identification method and system oriented to scientific literature knowledge entity
CN114943216B (en) Case microblog attribute level view mining method based on graph attention network
CN111708896B (en) Entity relationship extraction method applied to biomedical literature

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