Method for human-machine dialogue, computing device and computer-readable storage medium

Patent metadata

Field	Value
Publication number	US-12282747-B2
Application number	US-202217870813-A
Country	US
Kind code	B2
Filing date	Jul 21, 2022
Priority date	Dec 29, 2020
Publication date	Apr 22, 2025
Grant date	Apr 22, 2025

Abstract

A method includes: acquiring an input sentence in a first language in a current round of conversation; translating the input sentence in the first language to obtain an input sentence in a second language, according to dialogue contents in the first language and dialogue contents in the second language that have a mutual translation relationship with the dialogue contents in the first language in historical rounds of conversation; invoking a multi-round conversation generation model to parse the input sentence in the second language in the current round of conversation to generate an output sentence in the second language in the current round of conversation; translating the output sentence in the second language in the current round of conversation to obtain at least one candidate result in the first language; and determining an output sentence in the first language from the at least one candidate result in the first language.

First claim

What is claimed is:

1. A computer-implemented method for human-machine dialogue, comprising:
acquiring an input sentence in a first language in a current round of conversation;
translating, through a first sentence translation model, the input sentence in the first language in the current round of conversation to obtain an input sentence in a second language in the current round of conversation, according to dialogue contents in the first language and dialogue contents in the second language that have a mutual translation relationship with the dialogue contents in the first language in historical rounds of conversation, wherein the dialogue contents in the first language comprise input sentences in the first language and output sentences in the first language in corresponding rounds of conversation, and the dialogue contents in the second language comprise input sentences in the second language and output sentences in the second language in corresponding rounds of conversation, and wherein the first sentence translation model is obtained by performing training parallel corpora between the second language and the first language of multiple rounds of continuous conversation in open domain;
invoking a pre-stored multi-round conversation generation model to parse the input sentence in the second language in the current round of conversation, according to the dialogue contents in the second language in the historical rounds of conversation, to generate an output sentence in the second language in the current round of conversation, wherein the multi-round conversation generation module is obtained by perform training based on multi-round dialogue corpora in the second language;
translating, through a second sentence translation model, the output sentence in the second language in the current round of conversation, according to the dialogue contents in the first language and in the second language in the historical rounds of conversation, and the input sentence in the first language and the input sentence in the second language in the current round of conversation, to obtain at least one candidate result in the first language, wherein the second sentence translation model is obtained by performing training parallel corpora between the first language and the second language of multiple rounds of continuous conversation in the open domain; and
determining an output sentence in the first language in the current round of conversation from the at least one candidate result in the first language for output.

2. The method of claim 1 , wherein determining the output sentence in the first language in the current round of conversation from the at least one candidate result in the first language for output comprises:
for each of the at least one candidate result in the first language, invoking a pre-stored coherence evaluation model to calculate an expression coherence of the candidate result in the first language according to the dialogue contents in the first language in the historical rounds of conversation and the input sentence in the first language in the current round of conversation; and
selecting one of the at least one candidate result with a largest expressive coherence as the output sentence in the first language in the current round of conversation and outputting the output sentence in the first language in the current round of conversation.

3. The method of claim 2 , further comprising:
acquiring a plurality of valid dialogue corpus samples in the first language, wherein each of the valid dialogue corpus samples comprises dialogue contents in the first language in multiple consecutive rounds of conversation;
for each of the valid dialogue corpus samples, constructing a negative dialogue corpus sample corresponding to the valid dialogue corpus sample; and
training an initial classification model using the plurality of the valid dialogue corpus samples and the negative dialogue corpus samples to obtain the coherence evaluation model.

4. The method of claim 3 , wherein constructing a negative dialogue corpus sample corresponding to the valid dialogue corpus sample comprises:
extracting, from the plurality of valid dialogue corpus samples, a target output sentence in the first language corresponding to a last one of the multiple consecutive rounds of conversation;
translating the target output sentence in the first language to obtain a corresponding expression sentence in the second language;
translating the expression sentence in the second language to a corresponding expression sentence in the first language;
calculating a minimum edit distance between the target output sentence in the first language and the expression sentence in the first language;
comparing the minimum edit distance with a preset distance threshold to obtain a comparison result, and determining a negative output sentence in the first language matching the target output sentence in the first language according to the comparison result; and
replacing the target output sentence in the first language in valid dialogue corpus samples with the negative output sentence in the first language to obtain the negative dialogue corpus sample.

5. The method of claim 4 , wherein determining the negative output sentence in the first language matching the target output sentence in the first language according to the comparison result comprises:
in response to the minimum edit distance being greater than the preset distance threshold, determining the expression sentence in the first language as the negative output sentence in the first language; and
in response to the minimum edit distance being equal to the preset distance threshold, performing a synonym substitution on at least one word in the expression sentence in the first language to obtain the negative output sentence in the first language.

6. The method of claim 1 , wherein each parallel corpus required for training the first sentence translation model is expressed by a long corpus sequence as follows:
[ . . . ,L t-2 ,H t-2 ,L t-1 ,H t-1 ,L t ,H t ],
where, t is an integer, H represents sentences in the second language, L represents sentences in the first language, and H t-n represents a translated parallel corpus of L t-n that has same subscript with H t-n , wherein n=0, 1, 2, 3, 4 . . . ;
wherein a first-language conversation operation of a same round of conversation comprises: an input operation of an input sentence in the first language, and an output operation of an output sentence in the first language in a round of conversation; and
a second-language conversation operation of the same round of conversation comprises: an input operation of an input sentence in the second language, and an output operation of an output sentence in the second language in the round of conversation.

7. The method of claim 6 , wherein, when t is an even number, L t represents an output sentence in the first language in the t/2th round of conversation, L t-1 represents an input sentence in the first language in the t/2th round of conversation, H t represents an output sentence in the second language in the t/2th round of conversation, H t-1 represents an input sentence in the second language in the t/2th round of conversation, H t-2 represents an output sentence in the second language in the (t−2)/2th round of conversation, and L t-2 represents an output sentence in the first language in the (t−2)/2th round of conversation.

8. The method of claim 6 , wherein, when t is an odd number, L t represents an input sentence in the first language in the (t+1)/2th round of conversation, L t-1 represents an output sentence in the first language in the (t−1)/2th round of conversation, H t represents an input sentence in the second language in the (t+1)/2

Assignees

• Ubtech Robotics Corp Ltd

Inventors

• Ma Li

Classifications

• G06F16/3343

  using phonetics · CPC title

• G06N20/00

  Machine learning · CPC title

• G06F16/90332

  Natural language query formulation or dialogue systems · CPC title

• G06F16/3337

  Translation of the query language, e.g. Chinese to English · CPC title

• G06F40/47

  Machine-assisted translation, e.g. using translation memory · CPC title