# US 2019/0377794 A1 · Method and apparatus for determining user intent

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

- Number: US 2019/0377794 A1
- Document ID: US20190377794A1
- Status: Published application
- Kind: A1
- Application number: 16/352,159
- Filed: 2019-03-13
- Published: 2019-12-12
- Earliest claimed: 2018-06-07
- Assignees: Alibaba Group Holding Limited
- Applicants: Alibaba Group Holding Limited
- Inventors: Ling LI; Zhiwei SHI; Yanjie LIANG
- Web page: https://patentgrep.com/patent/US20190377794A1

## CPC

- G06F 40/30: Semantic analysis
- G06F 18/24: Classification techniques
- G06F 18/24323: {Tree-organised classifiers}
- G06F 40/216: using statistical methods
- G06V 10/40: Extraction of image or video features
- G06V 30/153: {using recognition of characters or words}
- G06V 30/274: {Syntactic or semantic context, e.g. balancing}
- G06V 30/40: Document-oriented image-based pattern recognition
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING

## Abstract

The disclosed embodiments describe methods, systems, and apparatuses for determining user intent. A method is disclosed comprising obtaining a session text of a user; calculating, by the processor, a feature vector based on the session text; determining probabilities that the session text belongs to a plurality of intent labels, the probabilities calculated using a multi-level hierarchal intent classification model, the intent labels assigned to levels in the multi-level hierarchal intent classification model; and assigning a user intent to the session text based on the probabilities.

## Claims

### Claim 1

A method comprising:
obtaining, by a processor, a session text of a user;
calculating, by the processor, a feature vector based on the session text;
determining, by the processor, probabilities that the session text belongs to a plurality of intent labels, the probabilities calculated using a multi-level hierarchal intent classification model, the intent labels assigned to levels in the multi-level hierarchal intent classification model; and
assigning, by the processor, a user intent to the session text based on the probabilities.

### Claim 2 (depends on 1)

The method of claim 1, the calculating the feature vector comprising:
identifying, by the processor, a word segment set corresponding to the session text, the word segment set comprising a plurality of word segments;
calculating, by the processor, a weight value for each of the word segments;
constructing, by the processor, a word vector for each of the word segments;
constructing, by the processor, a text vector corresponding to the session text based on the word vector of each of the word segments and the corresponding weight values; and
using, by the processor, the feature vector as the text vector.

### Claim 3 (depends on 2)

The method of claim 2, the constructing the text vector comprising:
performing, by the processor, a weighted summation on word vectors of each of the word segments according to the corresponding weight values of each word segment, the weighted summation generating a sum vector; and
using, by the processor, the sum vector as the corresponding text vector

### Claim 4 (depends on 2)

The method of claim 2, the using the feature vector as the text vector comprising:
combining, by the processor, the weight values of each of the word segments into a weight vector;
combining, by the processor, the weight vector and the text vector into a combined vector; and
using, by the processor, the combined vector as the feature vector.

### Claim 5 (depends on 1)

The method of claim 1, the determining probabilities that the session text belongs to the plurality of intent labels comprising:
inputting, by the processor, the feature vector into a plurality of classifiers corresponding to the levels of the multi-level hierarchal intent classification model; and
obtaining, by the processor, probabilities of the session text belonging to each of the intent labels in each of the levels.

### Claim 6 (depends on 1)

The method of claim 1, the determining probabilities that the session text belongs to the plurality of intent labels comprising:
inputting, by the processor, the feature vector into a first classifier corresponding to a first level of the levels;
obtaining, by the processor, probabilities of the session text belonging to each of the intent labels in the first level based on output of the first classifier;
determining, by the processor, a combination vector based on the feature vector and the probabilities of the session text belonging to each of the intent labels in the first level; and
inputting, by the processor, the combination vector into a second classifier corresponding to a second level of the multi-level hierarchal intent classification model, and obtaining probabilities of the session text belonging to each intent label in the second level.

### Claim 7 (depends on 1)

The method of claim 1, the assigning the user intent to the session text based on the probabilities comprising determining, by the processor, the user intent based on the probabilities and probability thresholds corresponding to each of the levels of the multi-level hierarchal intent classification model.

### Claim 8

A non-transitory computer readable storage medium for tangibly storing computer program instructions capable of being executed by a computer processor, the computer program instructions defining the steps of:
obtaining a session text of a user;
calculating a feature vector based on the session text;
determining probabilities that the session text belongs to a plurality of intent labels, the probabilities calculated using a multi-level hierarchal intent classification model, the intent labels assigned to levels in the multi-level hierarchal intent classification model; and
assigning a user intent to the session text based on the probabilities.

### Claim 9 (depends on 8)

The non-transitory computer readable storage medium of claim 8, the calculating the feature vector comprising:
identifying a word segment set corresponding to the session text, the word segment set comprising a plurality of word segments;
calculating a weight value for each of the word segments;
constructing a word vector for each of the word segments;
constructing a text vector corresponding to the session text based on the word vector of each of the word segments and the corresponding weight values; and
using the feature vector as the text vector.

### Claim 10 (depends on 9)

The non-transitory computer readable storage medium of claim 9, the constructing the text vector comprising:
performing a weighted summation on word vectors of each of the word segments according to the corresponding weight values of each word segment, the weighted summation generating a sum vector; and
using the sum vector as the corresponding text vector

### Claim 11 (depends on 9)

The non-transitory computer readable storage medium of claim 9, the using the feature vector as the text vector comprising:
combining the weight values of each of the word segments into a weight vector;
combining the weight vector and the text vector into a combined vector; and
using the combined vector as the feature vector.

### Claim 12 (depends on 8)

The non-transitory computer readable storage medium of claim 8, the determining probabilities that the session text belongs to the plurality of intent labels comprising:
inputting the feature vector into a plurality of classifiers corresponding to the levels of the multi-level hierarchal intent classification model; and
obtaining probabilities of the session text belonging to each of the intent labels in each of the levels.

### Claim 13 (depends on 8)

The non-transitory computer readable storage medium of claim 8, the determining probabilities that the session text belongs to the plurality of intent labels comprising:
inputting the feature vector into a first classifier corresponding to a first level of the levels;
obtaining probabilities of the session text belonging to each of the intent labels in the first level based on output of the first classifier;
determining a combination vector based on the feature vector and the probabilities of the session text belonging to each of the intent labels in the first level; and
inputting the combination vector into a second classifier corresponding to a second level of the multi-level hierarchal intent classification model, and obtaining probabilities of the session text belonging to each intent label in the second level.

### Claim 14 (depends on 8)

The non-transitory computer readable storage medium of claim 8, the assigning the user intent to the session text based on the probabilities comprising determining the user intent based on the probabilities and probability thresholds corresponding to each of the levels of the multi-level hierarchal intent classification model.

### Claim 15

An apparatus comprising:
a processor; and
a storage medium for tangibly storing thereon program logic for execution by the processor, the stored program logic comprising:
logic, executed by the processor, for obtaining a session text of a user,
logic, executed by the processor, for calculating a feature vector based on the session text,
logic, executed by the processor, for determining probabilities that the session text belongs to a plurality of intent labels, the probabilities calculated using a multi-level hierarchal intent classification model, the intent labels assigned to levels in the multi-level hierarchal intent classification model, and
logic, executed by the processor, for assigning a user intent to the session text based on the probabilities.

### Claim 16 (depends on 15)

The apparatus of claim 15, the logic for calculating the feature vector comprising:
logic, executed by the processor, for identifying a word segment set corresponding to the session text, the word segment set comprising a plurality of word segments;
logic, executed by the processor, for calculating a weight value for each of the word segments;
logic, executed by the processor, for constructing a word vector for each of the word segments;
logic, executed by the processor, for constructing a text vector corresponding to the session text based on the word vector of each of the word segments and the corresponding weight values; and
logic, executed by the processor, for using the feature vector as the text vector.

### Claim 17 (depends on 16)

The apparatus of claim 16, the logic for constructing the text vector comprising:
logic, executed by the processor, for performing a weighted summation on word vectors of each of the word segments according to the corresponding weight values of each word segment, the weighted summation generating a sum vector; and
logic, executed by the processor, for using the sum vector as the corresponding text vector

### Claim 18 (depends on 16)

The apparatus of claim 16, the logic for using the feature vector as the text vector comprising:
logic, executed by the processor, for combining the weight values of each of the word segments into a weight vector;
logic, executed by the processor, for combining the weight vector and the text vector into a combined vector; and
logic, executed by the processor, for using the combined vector as the feature vector.

### Claim 19 (depends on 15)

The apparatus of claim 15, the logic for determining probabilities that the session text belongs to the plurality of intent labels comprising:
logic, executed by the processor, for inputting the feature vector into a plurality of classifiers corresponding to the levels of the multi-level hierarchal intent classification model; and
logic, executed by the processor, for obtaining probabilities of the session text belonging to each of the intent labels in each of the levels.

### Claim 20 (depends on 15)

The apparatus of claim 15, the logic for determining probabilities that the session text belongs to the plurality of intent labels comprising:
logic, executed by the processor, for inputting the feature vector into a first classifier corresponding to a first level of the levels;
logic, executed by the processor, for obtaining probabilities of the session text belonging to each of the intent labels in the first level based on output of the first classifier;
logic, executed by the processor, for determining a combination vector based on the feature vector and the probabilities of the session text belonging to each of the intent labels in the first level; and
logic, executed by the processor, for inputting the combination vector into a second classifier corresponding to a second level of the multi-level hierarchal intent classification model, and obtaining probabilities of the session text belonging to each intent label in the second level.
