# US 2020/0394185 A1 · Intelligent knowledge-learning and question-answering

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

- Number: US 2020/0394185 A1
- Document ID: US20200394185A1
- Status: Published application
- Kind: A1
- Application number: 16/975,373
- Filed: 2019-02-27
- Published: 2020-12-17
- Earliest claimed: 2018-03-06
- Assignees: Microsoft Technology Licensing, LLC
- Applicants: Microsoft Technology Licensing, LLC
- Inventors: Yang Tian; Jianyong Wang; Peng Chen; Wei Wang; Ting Sun; Jie Zhang
- Web page: https://patentgrep.com/patent/US20200394185A1

## CPC

- G06F 16/2428: {Query predicate definition using graphical user interfaces, including menus and forms (G06F16/2423 takes precedence)}
- G06N 5/02: Knowledge representation; Symbolic representation
- G06F 16/3344: {using natural language analysis}
- G06F 40/211: Syntactic parsing, e.g. based on context-free grammar [CFG] or unification grammars
- G06F 40/284: Lexical analysis, e.g. tokenisation or collocates
- G06F 40/30: Semantic analysis
- G06F 40/35: Discourse or dialogue representation
- G06N 3/0499: Feedforward networks
- G06N 3/08: Learning methods
- G06N 3/09: Supervised learning
- G06N 5/022: Knowledge engineering; Knowledge acquisition
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

The present disclosure provides a technical solution related to intelligent knowledge learning and question-answering. The technical solution involves extracting knowledge from corpus and providing a user with a question-answering service based on the extracted knowledge.

## Claims

### Claim 1

A device, comprising:
an extracting module configured to extract an entity and a first predicate from a user's query;
a predicate searching module configured to search for a plurality of second predicates matched with the entity in a knowledge database according to the entity in the user's query;
a predicate selecting module configured to calculate correlations between the first predicate and the plurality of second predicates, and perform selection on the plurality of second predicates according to the correlations to determine one or more third predicates; and
a knowledge acquiring module configured to perform searching in the knowledge database based on the third predicate and the entity in the user's query to acquire knowledge.

### Claim 2 (depends on 1)

The device according to claim 1, wherein the predicate selecting module comprises:
a semantic relevance analyzing module configured to convert the first predicate and the plurality of second predicates into semantic vectors, and perform semantic relevance analyzing to generate a rank of correlations between the plurality of second predicates and the first predicate and/or correlations between the plurality of second predicates and the first predicate; and
a predicate determining module configured to determine one or more third predicates according to a preset correlation threshold and/or a correlation rank selecting strategy.

### Claim 3 (depends on 2)

The device according to claim 2, wherein the performing semantic relevance analyzing comprises: performing semantic relevance analyzing based on a neural network model.

### Claim 4 (depends on 1)

The device according to claim 1, wherein the knowledge is stored in the form of triple in the knowledge database, and the device further comprises:
a word-pair recognizing module configured to recognize one or more pairs of words correlated with each other from an input corpus, wherein at least one word in the pair of words is to be used as an entity in the triple;
an implied predicate determining module configured to recognize relationship between the words in the pair of words and determine an implied predicate in the pair of words according to a preset semantic rule of predicates; and
an implicit triple generating module configured to generate the implicit triple based on the implied predicate and the pair of words.

### Claim 5 (depends on 4)

The device according to claim 4, further comprising:
a word-pair selecting module configured to perform selection on the one or more pairs of words according to occurrence rates of the pairs of words, and provide the selected one or more pairs of words to the implied predicate determining module.

### Claim 6 (depends on 1)

The device according to claim 1, wherein the knowledge is stored in the form of triple in the knowledge database, and the device further comprises:
a word-vector converting module configured to extract entities and predicates from a plurality of triples in the knowledge database, and convert the entities and the predicates into word vectors;
a vector distance calculating module configured to calculate vector distances between the word vectors corresponding to the entities and the predicates in the triples; and
a triple selecting module configured to perform selection on the plurality of triples according to the vector distances.

### Claim 7 (depends on 6)

The device according to claim 6, wherein the converting the entities and the predicates into word vectors comprises:
performing word embedding based on neural network on the entities and the predicates in the plurality of triples based on a given corpus to generate word vectors.

### Claim 8

A method, comprising:
extracting an entity and a first predicate from a user's query;
searching for a plurality of second predicates matched with the entity in a knowledge database according to the entity in the user's query;
calculating correlations between the first predicate and the plurality of second predicates, and performing selection on the plurality of second predicates according to the correlations to determine one or more third predicates; and
performing searching in the knowledge database based on the third predicate and the entity in the user's query to acquire knowledge.

### Claim 9 (depends on 8)

The method according to claim 8, wherein the calculating correlations between the first predicate and the plurality of second predicates, and performing selection on the plurality of second predicates according to the correlations to determine one or more third predicates comprises:
converting the first predicate and the plurality of second predicates into semantic vectors, and performing semantic relevance analyzing to generate a rank of correlations between the plurality of second predicates and the first predicate and/or correlations between the plurality of second predicates and the first predicate; and
determining one or more third predicates according to a preset correlation threshold and/or a correlation rank selecting strategy.

### Claim 10 (depends on 9)

The method according to claim 9, wherein the performing semantic relevance analyzing comprises: performing semantic relevance analyzing based on a neural network model.

### Claim 11 (depends on 8)

The method according to claim 8, wherein the knowledge is stored in the form of triple in the knowledge database, and the method further comprises extracting implicit triples comprising:
recognizing one or more pairs of words correlated with each other from an input corpus, wherein at least one word in the pair of words is to be used as an entity in the triple;
recognizing relationship between the words in the pair of words and determining an implied predicate in the pair of words according to a preset semantic rule of predicates; and
generating the implicit triple based on the implied predicate and the pair of words.

### Claim 12 (depends on 11)

The method according to claim 11, after recognizing the pairs of words correlated with each other, the method further comprises:
performing selection on the one or more pairs of words according to occurrence rates of the pairs of words, and determining the implied predicate according to the selected one or more pairs of words.

### Claim 13 (depends on 8)

The method according to claim 8, wherein the knowledge is stored in the form of triple in the knowledge database, and the method further comprises removing noise from a plurality of extracted triples, and the removing noise from a plurality of extracted triples comprises:
extracting entities and predicates from a plurality of triples in the knowledge database, and converting the entities and the predicates into word vectors;
calculating vector distances between the word vectors corresponding to the entities and the predicates in the triples; and
performing selection on the plurality of triples according to the vector distances.

### Claim 14 (depends on 13)

The method according to claim 13, wherein the converting the entities and the predicates into word vectors comprises:
performing word embedding based on neural network on the entities and the predicates in the plurality of triples based on a given corpus to generate word vectors.

### Claim 15

An electronic apparatus, comprising:
a processing unit; and
a memory, coupled to the processing unit and containing instructions stored thereon, the instructions cause the electronic apparatus to perform operations upon being executed by the processing unit, the operations comprise:
extracting an entity and a first predicate from a user's query;
searching for a plurality of second predicates matched with the entity in a knowledge database according to the entity in the user's query;
calculating correlations between the first predicate and the plurality of second predicates, and performing selection on the plurality of second predicates according to the correlations to determine one or more third predicates; and
performing searching in the knowledge database based on the third predicate and the entity in the user's query to acquire knowledge.
