Method and apparatus for recognizing user intention, device, and readable storage medium

Patent metadata

Field	Value
Publication number	US-11646016-B2
Application number	US-202017024668-A
Country	US
Kind code	B2
Filing date	Sep 17, 2020
Priority date	Feb 19, 2020
Publication date	May 9, 2023
Grant date	May 9, 2023

Abstract

A method and apparatus for recognizing a user intention, a device, and a readable storage medium are disclosed. A specific implementation of the method includes: acquiring a plurality of rounds of dialogue data and a satisfaction degree of each round of dialogue data, selecting target dialogue data having a satisfaction degree meeting a set satisfaction condition; labeling input data in the target dialogue data with an intention label; and training the intention recognition model based on the input data in the target dialogue data and the intention label of the input data, so that the trained intention recognition model performs intention recognition on new input data.

First claim

What is claimed is:

1. A method for recognizing a user intention, comprising:
acquiring a plurality of rounds of dialogue data and a satisfaction degree of each round of dialogue data, each round of dialogue data comprising input data of a user and response data, the response data being fed back based on an intention recognition result obtained by an intention recognition model performing an intention recognition on the input data;
selecting, from the plurality of rounds of dialogue data, target dialogue data having a satisfaction degree meeting a set satisfaction condition;
labeling input data in the target dialogue data with an intention label;
acquiring types of source terminals of the input data in the target dialogue data; and
training the intention recognition model based on the input data in the target dialogue data and the intention label of the input data, so that the trained intention recognition model performs intention recognition on new input data, wherein the training the intention recognition model comprises:
grouping the input data in the target dialogue data according to the types of the source terminals to obtain a plurality of groups of input data; and
training, based on a group of input data and an intention label of the group of input data, an intention recognition model corresponding to a type of a source terminal of the group of input data, wherein, the types of the source terminals comprise a screen terminal type and a screenless terminal type.

2. The method according to claim 1 , wherein the training the intention recognition model based on the input data in the target dialogue data and the intention label of the input data, comprises:
extracting a linguistic feature of the input data in the target dialogue data and a feature associated with the input data; and
training the intention recognition model by using the linguistic feature of the input data and the feature associated with the input data as an input, and using the intention label as a target output,
wherein, the feature associated with the input data comprises at least one of: a search demand of the input data, response data of the input data, context data of the input data, or statistical data of the input data.

3. The method according to claim 1 , wherein the acquiring the plurality of rounds of dialogue data and the satisfaction degree of each round of dialogue data, comprises:
pulling, according to a set period, a satisfaction degree feedback log of a most recent period, the satisfaction degree feedback log comprising the plurality of rounds of dialogue data and the satisfaction degree of each round of dialogue data;
the training the intention recognition model based on the input data in the target dialogue data and the intention label of the input data, comprises:
training, according to the set period, the intention recognition model based on the input data in the target dialogue data and the intention label of the input data.

4. The method according to claim 3 , wherein, before the pulling the satisfaction degree feedback log of the most recent period according to the set period, the satisfaction degree feedback log comprising the plurality of rounds of dialogue data and the satisfaction degree of each round of dialogue data, the method further comprises:
acquiring the plurality of rounds of dialogue data from a user behavior log;
scoring, by using a satisfaction degree scoring model, the satisfaction degree of each round of dialogue data in the plurality of rounds of dialogue data, to obtain a satisfaction degree score of each round of dialogue data; and
storing the plurality of rounds of dialogue data and the satisfaction degree score of each round of dialogue data to the satisfaction degree feedback log.

5. The method according to claim 4 , wherein the acquiring the plurality of rounds of dialogue data from the user behavior log, comprises:
collecting a plurality of rounds of candidate dialogue data using the user behavior log, wherein each round of candidate dialogue data comprises input data of a user and response data, the response data being fed back based on an intention recognition result obtained by an intention recognition model performing an intention recognition on the input data;
acquiring, based on the intention recognition model, an intention confidence of the input data in each round of candidate dialogue data; and
selecting, from the user behavior log, input data having an intention confidence meeting a set confidence condition and response data of the input data, to form the plurality of rounds of dialogue data.

6. The method according to claim 1 , wherein the labeling the input data in the target dialogue data with the intention label, comprises:
acquiring category information and instruction information of the response data in the target dialogue data;
determining the intention label of the input data in the target dialogue data based on the category information and the instruction information, and a mapping relationship between the intention label and the category information and the instruction information; and
labeling the input data in the target dialogue data with the intention label.

7. An electronic device, comprising:
at least one processor; and
a memory, communicatively connected to the at least one processor; wherein,
the memory, storing instructions executable by the at least one processor, the instructions, when executed by the at least one processor, cause the at least one processor to perform operations, the operations comprising:
acquiring a plurality of rounds of dialogue data and a satisfaction degree of each round of dialogue data, each round of dialogue data comprising input data of a user and response data, the response data being fed back based on an intention recognition result obtained by an intention recognition model performing an intention recognition on the input data;
selecting, from the plurality of rounds of dialogue data, target dialogue data having a satisfaction degree meeting a set satisfaction condition;
labeling input data in the target dialogue data with an intention label;
acquiring types of source terminals of the input data in the target dialogue data; and
training the intention recognition model based on the input data in the target dialogue data and the intention label of the input data, so that the trained intention recognition model performs intention recognition on new input data, wherein the training the intention recognition model comprises:
grouping the input data in the target dialogue data according to the types of the source terminals to obtain a plurality of groups of input data; and
training, based on a group of input data and an intention label of the group of input data, an intention recognition model corresponding to a type of a source terminal of the group of input data, wherein, the types of the source terminals comprise a screen terminal type and a screenless terminal type.

8. The electronic device according to claim 7 , wherein the training the intention recognition model based on the input data in the target dialogue data and the intention label of the input data, comprises:
extracting a linguistic feature of the input data in the target dialogue data and a feature associated with the input data; and
training the intention recognition model by using the linguistic feature of the input data and the feature associated with the input data as an input, and using the intention label as a target output,
wherein, the feature associated with the input data comprises at least one of: a search demand of the input data, response data of the input data, context data of the input data, or statistical data of the input data.

9. The elect

Assignees

• Baidu online network technology beijing co ltd
• Shanghai Xiaodu Tech Co Ltd

Inventors

• Zhai Zhouwei
• Huang Rongsheng
• Xu Ben
• You Fei
• Li Ning
• He Wensong

Classifications

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N5/04

  Inference or reasoning models · CPC title

• G06N20/00

  Machine learning · CPC title