# US 2019/0171642 A1 · Method and system for automatically discovering inference rule, database and retrieval method

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2019/0171642 A1
- Document ID: US20190171642A1
- Status: Published application
- Kind: A1
- Application number: 16/317,686
- Filed: 2018-01-17
- Published: 2019-06-06
- Earliest claimed: 2017-06-28
- Assignees: Boe Technology Group Co., Ltd.
- Applicants: Boe Technology Group Co., Ltd.
- Inventors: Zhenzhong ZHANG
- Web page: https://patentgrep.com/patent/US20190171642A1

## CPC

- G06F 16/2465: {Query processing support for facilitating data mining operations in structured databases}
- G16H 70/00: ICT specially adapted for the handling or processing of medical references
- G06F 16/288: {Entity relationship models}
- G06N 5/04: Inference or reasoning models
- G16H 50/70: for mining of medical data, e.g. analysing previous cases of other patients
- G06F 17/16: Matrix or vector computation {, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization (matrix transposition G06F7/78)}
- G06F 2216/03: Data mining
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G16H: HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

The disclosure provides a method and a system for automatically discovering an inference rule, a medical knowledge database and a retrieval method. The method includes: acquiring a relation matrix for each medical relation, the relation matrix representing a relationship between vectors of the medical entities having the medical relation; and acquiring the inference rules based on the relation matrices for the plurality of medical relations, each inference rule including a first medical relation, a second medical relation and a third medical relation, the inference rule representing that the third medical relation is inferred from the first medical relation and the second medical relation.

## Claims

### Claim 1

A computing device implemented method for providing knowledge relations in a knowledge base, the knowledge base comprising a plurality of knowledges, each knowledge comprising two knowledge entities and one knowledge relation between the two knowledge entities, the method comprising:
determining a relation matrix for each knowledge relation using one or more processors, the relation matrix representing a relationship between vectors of the knowledge entities having the knowledge relation;
acquiring the inference rules results in accordance with the relation matrices for the plurality of knowledge relations using one or more processors, each inference result comprising a first knowledge relation, a second knowledge relation and a third knowledge relation, the inference results representing that the third knowledge relation is inferred from the first knowledge relation and the second knowledge relation.

### Claim 2 (depends on 1)

The computing device implemented method according to claim 1, wherein the acquiring the relation matrix for each knowledge relation comprises:
providing an initial relation matrix for each knowledge relation;
acquiring a knowledge corresponding to the knowledge relation in the knowledge base as a correct knowledge;
changing a knowledge entity in the correct knowledge to acquire a new knowledge as a wrong knowledge;
scoring the correct knowledge and the wrong knowledge using a scoring function respectively; and
performing a learning operation on the initial relation matrix using a predetermined target function so that a score of the correct knowledge is higher than a score of the wrong knowledge, and acquiring the relation matrix for the knowledge relation.

### Claim 3 (depends on 2)

The computing device implemented method according to claim 2, wherein a difference between the number of the correct knowledges and the number of the wrong knowledges is smaller than a predetermined threshold when the learning operation is performed on the initial relation matrix.

### Claim 4 (depends on 2)

The computing device implemented method according to claim 2, wherein the changing the knowledge entity in the correct knowledge so as to acquire the new knowledge as the wrong knowledge comprises replacing the knowledge entity in the correct knowledge with a knowledge entity other than knowledge entities in all the correct knowledges corresponding to the knowledge relation, so as to acquire the new knowledge as the wrong knowledge.

### Claim 5 (depends on 2)

The computing device implemented method according to claim 2, wherein prior to acquiring the relation matrix for each knowledge relation, the method further comprises:
acquiring a one-hot code vector for each knowledge entity through the following formula:
x
e
=
(
a
1
,
…

,
a
i
,
…

,
a
N
)
,
a
j
=
{
1
,
if


j
=
i
0
,
if


j
≠
i
,
wherein N represents the number of the knowledge entities in the knowledge base, and xe represents a vector in N*1 dimensions; and
acquiring a vector of each knowledge in accordance with the one-hot code vector for each knowledge entity through the following formula:
ye=σ(Wxe), σ(x)=1/(1+ex)
wherein ye represents a vector in K*1 dimensions, W represents a mapping matrix and is a vector in K*N dimensions, and K is a predetermined value smaller than N.

### Claim 6 (depends on 5)

The computing device implemented method according to claim 5, wherein the scoring function is:
Score(r(e1,e2))=ye1TMrye2
where r represents the knowledge relation, e1 and e2 represent the knowledge entities, r(e1, e2) represents the knowledge, Score(·) represents the scoring function, and Mr represents the relation matrix for the knowledge relation r.

### Claim 7 (depends on 6)

The computing device implemented method according to claim 6, wherein the predetermined target function is:
L
=
∑
r

(
e
1
,
e
2
)
∈
T

∑
r

(
e
1
′
,
e
2
′
)
∈
T
′

max

{
Score

(
r

(
e
1
′
,
e
2
′
)
)
-
Score

(
r

(
e
1
,
e
2
)
)
+
1
,
0
}
where L represents the target function, T represents a set of the correct knowledges, T′ represents a set of the wrong knowledges, and Mr represents the relation matrix for the knowledge relation r.

### Claim 8 (depends on 1)

The computing device implemented method according to claim 1, wherein the acquiring the inference result in accordance with the relation matrix for the knowledge relation comprises:
selecting a plurality of groups of knowledge relations from all the knowledge relations, and creating a plurality of to-be-validated inference results, each of the plurality of to-be-validated inference results corresponding to one group of knowledge relations, each group of knowledge relations including three knowledge relations;
calculating a similarity between a product of relation matrices for two knowledge relations in each group of knowledge relations and a relation matrix for the other knowledge relation in the group of knowledge relations; and
selecting parts of the to-be-validated inference results as the final inference results in accordance with the similarities.

### Claim 9 (depends on 8)

The computing device implemented method according to claim 8, wherein the selecting the parts of the to-be-validated inference results as the final inference results in accordance with the similarities comprises:
ranking the similarities corresponding to all the groups of knowledge relations and obtaining a predetermined number of to-be-validated inference results having the largest similarity; and
determining the predetermined number of to-be-validated inference results having the largest similarity as the final inference results.

### Claim 10 (depends on 8)

The computing device implemented method according to claim 8, wherein the three knowledge relations in each group of knowledge relations meet the following conditions:
Lp∩Hq≠Ø, (1)
Hp∩Hr≠Ø, and (2)
Lq∩Lr≠Ø, (3)
where Hp, Hq and Hr represent sets of all first knowledge entities in triplets corresponding to the three knowledge relations p, q and r in the group of knowledge relations respectively, and Lp, Lq and Lr represent sets of all second knowledge entities in the triplets corresponding to the three knowledge relations p, q and r in the group of knowledge relations respectively.

### Claim 11 (depends on 8)

The computing device implemented method according to claim 8, wherein the calculating the similarity between the product of the relation matrices for two knowledge relations in each group of knowledge relations and the relation matrix for the other knowledge relation in the group of knowledge relations comprises calculating an L2 norm between the product of the relation matrices for two knowledge relations in each group of knowledge relations and the relation matrix for the other knowledge relation in the group of knowledge relations as the similarity.

### Claim 12

A computing device implemented system for providing knowledge relations in a knowledge base, the knowledge base comprising a plurality of knowledges, each knowledge comprising two knowledge entities and one knowledge relation, the system comprising:
a relation matrix learner configured to determine a relation matrix for each knowledge relation, the relation matrix representing a relationship between vectors of the knowledge entities having the knowledge relation; and
an inference result finder configured to acquire the inference results in accordance with the relation matrices for the plurality of knowledge relations, each inference results comprising a first knowledge relation, a second knowledge relation and a third knowledge relation, the inference results representing that the third knowledge relation is inferred from the first knowledge relation and the second knowledge relation.

### Claim 13 (depends on 12)

The computing device implemented system according to claim 12, wherein the relation matrix learner further comprises:
a providing circuit configured to provide an initial relation matrix for each knowledge relation;
a first acquiring circuit configured to acquire a knowledge corresponding to the knowledge relation in the knowledge base as a correct knowledge;
a changing circuit configured to change a knowledge entity in the correct knowledge so as to acquire a new knowledge as a wrong knowledge;
a scoring circuit configured to score the correct knowledge and the wrong knowledge using a scoring function respectively; and
a second acquiring circuit configured to acquire the relation matrix for the knowledge relation by performing a learning operation on the initial relation matrix using a predetermined target function to enable a score of the correct knowledge higher than a score of the wrong knowledge.

### Claim 14 (depends on 13)

The computing device implemented system according to claim 13, wherein the relation matrix learner comprises a calculating circuit configured to make a difference between the number of the correct knowledge and the number of the wrong knowledge smaller than a predetermined threshold when the learning operation is performed on the initial relation matrix.

### Claim 15 (depends on 13)

The computing device implemented system according to claim 13, wherein the relation matrix learner comprises a third acquiring circuit configured to acquire the new knowledge as the wrong knowledge by replacing the knowledge entity in the correct knowledge with a knowledge entity other than knowledge entities in all the correct knowledges corresponding to the knowledge relation.

### Claim 16 (depends on 13)

The computing device implemented system according to claim 13, further comprising an entity vector learner configured to acquire a one-hot code vector for each knowledge entity through the following formula:
x
e
=
(
a
1
,
…

,
a
i
,
…

,
a
N
)
,
a
j
=
{
1
,
if


j
=
i
0
,
if


j
≠
i
where N represents the number of the knowledge entities in the knowledge base, and xe represents a vector in N*1 dimensions; and acquire a vector of each knowledge entity in accordance with the one-hot code vector for each the knowledge entity through the following formula:
ye=σ(Wxe), σ(x)=1/(1+ex)
where ye represents a vector in K*1 dimensions, W represents a mapping matrix and is a vector in K*N dimensions, and K is a predetermined value smaller than N,
wherein the scoring function is:
Score(r(e1,e2))=ye1TMrye2
where r represents the knowledge relation, e1 and e2 represent the knowledge entities, r(e1, e2) represents the knowledge, Score(·) represents the scoring function, and Mr represents the relation matrix for the knowledge relation r,
wherein the predetermined target function is:
L
=
∑
r

(
e
1
,
e
2
)
∈
T

∑
r

(
e
1
′
,
e
2
′
)
∈
T
′

max

{
Score

(
r

(
e
1
′
,
e
2
′
)
)
-
Score

(
r

(
e
1
,
e
2
)
)
+
1
,
0
}
where L represents the target function, T represents a set of the correct knowledges, T′ represents a set of the wrong knowledges, and Mr represents the relation matrix for the knowledge relation r.

### Claim at position 17

17-22. (canceled)

### Claim 23 (depends on 1)

A data retrieval method from a knowledge base, comprising:
acquiring knowledge relations according to the method of claim 1;
inputting a search word or search pattern;
performing a search operation in accordance with the knowledge relations using the search word or search pattern, so as to acquire the knowledge related to the search word or search pattern; and
outputting the knowledge related to the search word or search pattern.

### Claim 24

(canceled)

### Claim 25

A knowledge database, comprises:
a data inputting device configured to input knowledge data;
a relation matrix learner configured to acquire a relation matrix for each knowledge relation, the relation matrix representing a relationship between vectors of knowledge entities having the knowledge relation, the knowledge base comprising a plurality of knowledges, each knowledge comprising two knowledge entities and one knowledge relation between the two knowledge entities;
an inference result finder configured to acquire the inference results in accordance with the relation matrices for the plurality of knowledge relations, each inference result comprising a first knowledge relation, a knowledge relation and a third knowledge medical relation, the inference result representing that the third knowledge relation is inferred from the first knowledge relation and the second knowledge relation;
an inputting device configured to input a search word or search pattern;
a searcher configured to perform a search operation in accordance with the inference result using the search word or search pattern, so as to acquire the knowledge related to the search word or search pattern;
an outputting device configured to output the knowledge related to the search word or search pattern; and
optionally an entity vector learner configured to acquire a one-hot code vector for each knowledge entity through the following formula:
x
e
=
(
a
1
,
…

,
a
i
,
…

,
a
N
)
,
a
j
=
{
1
,
if


j
=
i
0
,
if


j
≠
i
where N represents the number of the knowledge entities in the knowledge base, and xe represents a vector in N*1 dimensions; and acquire a vector of each knowledge entity in accordance with the one-hot code vector for each knowledge entity through the following formula:
ye=σ(Wxe), σ(x)=1/(1+ex)
where ye represents a vector in K*1 dimensions, W represents a mapping matrix and it is a vector in K*N dimensions, and K is a predetermined value smaller than N.

### Claim at position 21

26-27. (canceled)

### Claim 28 (depends on 1)

The computing device implemented method according to claim 1, wherein the plurality of knowledges are knowledges, the knowledge relations are medical relation, and the knowledge entities are medical entities.

### Claim 29 (depends on 12)

The computing device implemented system according to claim 12, wherein the plurality of knowledges are knowledges, the knowledge relations are medical relation, and the knowledge entities are medical entities.
