DYNAMIC SEMANTIC NETWORKS FOR LANGUAGE UNDERSTANDING AND QUESTION ANSWERING

• US 20190005090A1
• Filed: 06/29/2017
• Published: 01/03/2019
• Est. Priority Date: 06/29/2017
• Status: Active Grant

• Current Assignee: Futurewei Technologies Incorporated (Huawei Investment & Holding Co., Ltd.)
• Original Assignee: Futurewei Technologies Incorporated (Huawei Investment & Holding Co., Ltd.)
• Inventors: Zhang, Jun, Liu, Lifeng, Zhu, Yingxuan, Yin, Xiaotian, Li, Jian
• Granted Patent: US 11,243,944 B2
• Time in Patent Office: Days
• CPC Class Codes: G06F 16/2237 Vectors, bitmaps or matrices
• CPC Class Codes: G06F 16/2425 Iterative querying; Query formulation based on the results of a preceding query
• CPC Class Codes: G06F 16/3329 Natural language query formulation or dialogue systems
• CPC Class Codes: G06F 16/3338 Query expansion
• CPC Class Codes: G06F 40/30 Semantic analysis

Abstract

A computer-implemented method of answering questions comprises: receiving, by one or more processors, a query; based on the query, generating, by the one or more processors, a matrix; based on the matrix, modifying, by the one or more processors, a dynamic memory; based on the matrix, determining, by the one or more processors, a first response from the dynamic memory; based on the matrix, determining, by the one or more processors, a second response from a database; based on the first response and the second response, determining, by the one or more processors, a third response; and in response to the query, providing, by the one or more processors, the third response.

Claims

1. A computer-implemented method of answering questions, comprising:

receiving, by one or more processors, a query;

based on the query, generating, by the one or more processors, a matrix;

based on the matrix, modifying, by the one or more processors, a dynamic memory;

based on the matrix, determining, by the one or more processors, a first response from the dynamic memory;

based on the matrix, determining, by the one or more processors, a second response from a database;

based on the first response and the second response, determining, by the one or more processors, a third response; and

in response to the query, providing, by the one or more processors, the third response.

2. The computer-implemented method of claim 1, wherein:

the query comprises a plurality of words; and

the generating of the matrix comprises determining a vector for each word of the plurality of words.

3. The computer-implemented method of claim 2, wherein:

the vector for each word of the plurality of words has at least one hundred dimensions; and

the determining of the vector for each word of the plurality of words comprises retrieving a predetermined corresponding vector for the word from a dictionary database.

4. The computer-implemented method of claim 1, wherein:

the determining of the first response from the dynamic memory comprises;

providing the matrix as input to a long short-term memory (LSTM); and

in response to the input, receiving the first response from the LSTM.

5. The computer-implemented method of claim 1, wherein:

the method further comprises;

determining an interpolation gate factor using a training set; and

the determining of the third response is further based on the interpolation gate factor.

6. The computer-implemented method of claim 1, further comprising:

accessing an initial database that stores representations of facts as triplets, each triplet comprising a head entity, a relation, and a tail entity; and

generating the database from the initial database by performing operations comprising, for each triplet in the initial database;

generating a first vector based on the head entity and the relation;

generating a second vector based on the tail entity; and

storing the first vector and the second vector as a key-value pair, wherein the first vector is the key and the second vector is the value.

7. The computer-implemented method of claim 6, wherein the determining of the second response from the database comprises determining a cosine similarity between the matrix and one or more keys of the database.

8. The computer-implemented method of claim 7, wherein the determining of the second response from the database further comprises comparing each determined cosine similarity to a predetermined threshold.

9. A question-answering system comprising:

a memory storage comprising instructions; and

one or more processors in communication with the memory storage, wherein the one or more processors execute the instructions to perform;

receiving a query;

based on the query, generating a matrix;

based on the matrix, modifying a dynamic memory;

based on the matrix, determining a first response from the dynamic memory;

based on the matrix, determining a second response from a database;

based on the first response and the second response, determining a third response; and

in response to the query, providing the third response.

10. The system of claim 9, wherein:

the query comprises a plurality of words; and

the generating of the matrix comprises determining a vector for each word of the plurality of words.

11. The system of claim 10, wherein:

the vector for each word of the plurality of words has at least one hundred dimensions; and

the determining of the vector for each word of the plurality of words comprises retrieving a predetermined corresponding vector for the word from a dictionary database.

12. The system of claim 9, wherein:

the determining of the first response from the dynamic memory comprises;

providing the matrix as input to a long short-term memory (LSTM); and

in response to the input, receiving the first response from the LSTM.

13. The system of claim 9, wherein:

the one or more processors further perform;

determining an interpolation gate factor using a training set; and

the determining of the third response is further based on the interpolation gate factor.

14. The system of claim 9, wherein the one or more processors further perform:

accessing an initial database that stores representations of facts as triplets, each triplet comprising a head entity, a relation, and a tail entity; and

generating the database from the initial database by performing operations comprising, for each triplet in the initial database;

generating a first vector based on the head entity and the relation;

generating a second vector based on the tail entity; and

storing the first vector and the second vector as a key-value pair, wherein the first vector is the key and the second vector is the value.

15. The system of claim 14, wherein the determining of the second response from the database comprises determining a cosine similarity between the matrix and one or more keys of the database.

16. The system of claim 15, wherein the determining of the second response from the database further comprises comparing each determined cosine similarity to a predetermined threshold.

17. The system of claim 9, wherein:

the modifying of the dynamic memory comprises applying a weighting function; and

the one or more processors further perform;

determining a next value of the weighting function by convolving the weighting function with a shifting distribution.

18. A non-transitory computer-readable medium storing computer instructions for answering questions, that when executed by one or more processors, cause the one or more processors to perform steps of:

receiving a query;

based on the query, generating a matrix;

based on the matrix, modifying a dynamic memory;

based on the matrix, determining a first response from the dynamic memory;

based on the matrix, determining a second response from a database;

based on the first response and the second response, determining a third response; and

in response to the query, providing the third response.

19. The non-transitory computer-readable medium of claim 18, wherein:

the query comprises a plurality of words; and

the generating of the matrix comprises determining a vector for each word of the plurality of words.

20. The non-transitory computer-readable medium of claim 19, wherein:

the vector for each word of the plurality of words has at least one hundred dimensions; and

the determining of the vector for each word of the plurality of words comprises retrieving a predetermined corresponding vector for the word from a dictionary database.