Multi-document summary generation method and apparatus, and terminal

1. A multi-document summary generation method, comprising:

obtaining a candidate sentence set comprising a plurality of candidate sentences included in each candidate document of a plurality of candidate documents about an event;

processing each of the candidate sentences using a cascaded attention mechanism and an unsupervised learning model in a preset network model to obtain an importance of each of the candidate sentences, wherein an importance of a candidate sentence corresponds to a modulus of a row vector in a cascaded attention mechanism matrix, wherein the preset network model optimizes, based on each of the candidate sentences, m vectors describing the event, and a candidate matrix, a reconstruction error function from the unsupervised learning model to output the cascaded attention mechanism matrix, wherein the importance of the candidate sentence indicates an importance degree of a meaning from the candidate sentence in the candidate documents, wherein the reconstruction error function comprises a relationship between each of the candidate sentences, the m vectors, the candidate matrix, and a weight corresponding to the candidate matrix, wherein the candidate matrix is an m×n matrix, wherein m and n are positive integers, and wherein n is a quantity of words comprised in the candidate documents;

selecting, from the candidate sentence set based on the importance of each of the candidate sentences, a phrase that meets a preset condition as a summary phrase set; and

obtaining a summary of the candidate documents based on the summary phrase set.

2. The multi-document summary generation method of claim 1 , wherein the optimizing further comprises:

executing the unsupervised learning model; and

using the modulus of a row vector from the cascaded attention mechanism matrix to obtain the importance of each of the candidate sentences, and wherein the modulus of a row vector is the importance of the candidate sentence in a case of a minimum value of the reconstruction error function.

3. The multi-document summary generation method of claim 1 , wherein the selecting comprises:

filtering out, from the candidate sentences, a word that does not meet a preset rule in each of the candidate sentences to obtain each filtered candidate sentence;

extracting a first part-of-speech phrase and a second part-of-speech phrase from a syntax tree of each of the filtered candidate sentences to constitute a phrase set;

calculating, based on an importance of each of the filtered candidate sentences, an importance of the first part-of-speech phrase and an importance of the second part-of-speech phrase; and

selecting, from the phrase set and based on the importance of the first part-of-speech phrase and the importance of the second part-of-speech phrase, the first part-of-speech phrase and the second part-of-speech phrase that meet the preset condition as the summary phrase set.

4. The multi-document summary generation method of claim 3 , wherein the filtering comprises:

filtering out noise in each of the candidate sentences to obtain a candidate word set corresponding to each of the candidate sentences, wherein each of the candidate sentences comprises a plurality of words, and wherein a plurality of importance corresponds to the words; and

filtering out, based on the importance of each of the words, a word whose importance is less than a preset threshold in the candidate word set to obtain each of the filtered candidate sentences.

5. The multi-document summary generation method of claim 4 , wherein before filtering out the word whose importance is less than the preset threshold in the candidate word set, the method further comprises processing each of the candidate sentences using the cascaded attention mechanism and the unsupervised learning model in the preset network model to obtain an importance of each of a plurality of different words, wherein the candidate documents comprise the different words.

6. The multi-document summary generation method of claim 5 , wherein processing each of the candidate sentences using the cascaded attention mechanism and the unsupervised learning model in the preset network model to obtain the importance of the different words comprises optimizing, based on each of the candidate sentences, the m vectors used to describe the event, a candidate matrix, and the reconstruction error function, and wherein the optimizing further comprises:

executing the unsupervised learning model; and

using a modulus of a column vector of each column in the candidate matrix as the importance of the words when the reconstruction error function has a minimum value.

7. The multi-document summary generation method of claim 3 , wherein calculating, based on the importance of the candidate sentence, the importance of the first part-of-speech phrase and the importance of the second part-of-speech phrase that are extracted from each of the candidate sentences comprises:

obtaining a word frequency of the first part-of-speech phrase and the second part-of-speech phrase; and

calculating, based on the word frequency of the first part-of-speech phrase, the second part-of-speech phrase and the importance of the candidate sentence in which the first part-of-speech phrase and the second part-of-speech phrase are located, the importance of the first part-of-speech phrase and the importance of the second part-of-speech phrase.

8. The multi-document summary generation method of claim 3 , wherein selecting from the phrase set, based on the importance of the first part-of-speech phrase and the second part-of-speech phrase that correspond to each of the filtered candidate sentences, the first part-of-speech phrase and the second part-of-speech phrase that meet the preset condition as the summary phrase set, comprises:

inputting the importance of the first part-of-speech phrase and the second part-of-speech phrase and similarity between the first part-of-speech phrase and the second part-of-speech phrase into an integer linear programming function;

determining a candidate weight of the first part-of-speech phrase and the second part-of-speech phrase and a correlation weight of the similarity between the first part-of-speech phrase and the second part-of-speech phrase when an extremum is taken for the integer linear programming function, wherein the candidate weight of the first part-of-speech phrase or the second part-of-speech phrase is used to determine whether the first part-of-speech phrase or the second part-of-speech phrase meets the preset condition, wherein the correlation weight is used to determine whether similar phrases are simultaneously selected; and

determining, based on the candidate weight of the first part-of-speech phrase and the second part-of-speech phrase and the correlation weight of the similarity between the first part-of-speech phrase and the second part-of-speech phrase, a third part-of-speech phrase that meets the preset condition.

9. A multi-document summary generation apparatus, comprising:

a memory configured to store computer instructions; and

a processor coupled to the memory and configured to execute the computer instructions, which cause the multi-document summary generation apparatus to be configured to:

obtain a candidate sentence set comprising a plurality of candidate sentences included in each candidate document of a plurality of candidate documents about an event;

process each of the candidate sentences in the candidate sentence set using a cascaded attention mechanism and an unsupervised learning model in a preset network model to obtain the importance of each of the candidate sentences, wherein the importance of a candidate sentence corresponds to a modulus of a row vector in a cascaded attention mechanism matrix, wherein the preset network model optimizes, based on each of the candidate sentences, m vectors describing the event, and a candidate matrix, a reconstruction error function from the unsupervised learning model to output the cascaded attention mechanism matrix, wherein the importance of the candidate sentence indicates an importance degree of a meaning from the candidate sentence in the candidate documents, wherein the reconstruction error function comprises a relationship between each of the candidate sentences, the m vectors describing the event, the candidate matrix, and a weight corresponding to the candidate matrix, wherein the candidate matrix is an m×n matrix, wherein m and n are positive integers, and wherein n is a quantity of words comprised in the candidate documents;

select, from the candidate sentence set based on the importance of each of the candidate sentences, a phrase that meets a preset condition as a summary phrase set; and

obtain a summary of the candidate documents based on the summary phrase set.

10. The multi-document summary generation apparatus of claim 9 , wherein the computer instructions further cause the multi-document summary generation apparatus to be configured to:

execute the unsupervised learning model; and

use the modulus of a row vector from the cascaded attention mechanism matrix to obtain the importance of each of the candidate sentences, wherein the modulus of a row vector is the importance of the candidate sentence when the reconstruction error function is a minimum value.

11. The multi-document summary generation apparatus of claim 9 , wherein the computer instructions further cause the multi-document summary generation apparatus to be configured to:

filter out a word that does not meet a preset rule in each of the candidate sentences to obtain each filtered candidate sentence;

extract a first part-of-speech phrase and a second part-of-speech phrase from a syntax tree of each of the filtered candidate sentences to constitute a phrase set,

calculate, based on the importance of each of the filtered candidate sentences, an importance of the first part-of-speech phrase and the second part-of-speech phrase that are extracted from each of the filtered candidate sentences; and

select, from the phrase set based on the importance of the first part-of-speech phrase and the second part-of-speech phrase that correspond to each of the filtered candidate sentences, the first part-of-speech phrase and the second part-of-speech phrase that meet the preset condition as the summary phrase set.

12. The multi-document summary generation apparatus of claim 11 , wherein the computer instructions further cause the multi-document summary generation apparatus to be configured to:

filter out noise in each of the candidate sentences to obtain a candidate word set corresponding to each of the candidate sentences, wherein each of the candidate sentences comprises a plurality of words, and wherein a plurality of importance corresponds to the words; and

filter out, based on an importance of each of the words, a word whose importance is less than a preset threshold in the candidate word set to obtain each of the filtered candidate sentences.

13. The multi-document summary generation apparatus of claim 12 , wherein the computer instructions further cause the multi-document summary generation apparatus to be configured to train process each of the candidate sentences using the cascaded attention mechanism and the unsupervised learning model in the preset network model to obtain the importance of each of a plurality of different words, and wherein the candidate documents comprise the different words.

14. The multi-document summary generation apparatus of claim 13 , wherein the computer instructions further cause the multi-document summary generation apparatus to be configured to optimize, based on each of the candidate sentences, the m vectors used to describe the event, and a candidate matrix, the reconstruction error function, comprising:

executing the unsupervised learning model; and

using a modulus of a column vector of each column in the candidate matrix as the importance of the words when the reconstruction error function has a minimum value.

15. The multi-document summary generation apparatus of claim 11 , wherein the computer instructions further cause the multi-document summary generation apparatus to be configured to:

obtain a word frequency of the first part-of-speech phrase and the second part-of-speech phrase; and

calculate, based on the word frequency of the first part-of-speech phrase, the second part-of-speech phrase and the importance of the candidate sentence in which the first part-of-speech phrase and the second part of speech phrase are located, the importance of the first part-of-speech phrase and the second part-of-speech phrase .

16. The multi-document summary generation apparatus of claim 15 , wherein the computer instructions further cause the multi-document summary generation apparatus to be configured to:

input the importance of the first part-of-speech phrase and the second part-of-speech phrase and similarity between the first part-of-speech phrase and the second part-of-speech phrase into an integer linear programming function;

determine a candidate weight of the first part-of-speech phrase and the second part-of-speech phrase and a correlation weight of the similarity between the first part-of-speech phrase and the second part-of-speech phrase when an extremum is taken for the integer linear programming function, wherein the candidate weight of the first part-of-speech phrase and the second part-of-speech phrase is used to determine whether the first part-of-speech phrase and the second part-of-speech phrase meet the preset condition, wherein the correlation weight is used to determine whether similar phrases are simultaneously selected; and

determine, based on the candidate weight of the first part-of-speech phrase and the second part-of-speech phrase and the correlation weight of the similarity between the first part-of-speech phrase and the second part-of-speech phrase, a third part-of-speech phrase that meets the preset condition.

17. A computer program product comprising computer-executable instructions stored on a non-transitory computer-readable medium that, when executed by a processor, cause an apparatus to:

obtain a candidate sentence set comprising a plurality of candidate sentences included in each candidate document of a plurality of candidate documents about an event;

process each of the candidate sentences using a cascaded attention mechanism and an unsupervised learning model in a preset network model to obtain an importance of each of the candidate sentences, wherein an importance of a candidate sentence corresponds to a modulus of a row vector in a cascaded attention mechanism matrix, wherein the preset network model optimizes, based on each of the candidate sentences, m vectors describing the event, and a candidate matrix, a reconstruction error function from the unsupervised learning model to output the cascaded attention mechanism matrix, wherein the importance of the candidate sentence indicates an importance degree of a meaning from the candidate sentence in the candidate documents, wherein the reconstruction error function comprises a relationship between each of the candidate sentences, m vectors describing the event, a candidate matrix, and a weight corresponding to the candidate matrix, wherein the candidate matrix is an m×n matrix, wherein m and n are positive integers, and wherein n is a quantity of words comprised in the candidate documents;

select, from the candidate sentence set based on the importance of each of the candidate sentences, a phrase that meets a preset condition as a summary phrase set; and

obtain a summary of the candidate documents based on the summary phrase set.

18. The computer program product of claim 17 , wherein the computer-executable instructions further cause the apparatus to:

execute the unsupervised learning model; and

use the modulus of a row vector from the cascaded attention mechanism matrix to obtain the importance of each of the candidate sentences, wherein the modulus of a row vector is the importance of the candidate sentence when the reconstruction error function is a minimum value.

19. The computer program product of claim 17 , wherein the computer-executable instructions further cause the apparatus to:

filter out, from the candidate sentences, a word that does not meet a preset rule in each of the candidate sentences to obtain each filtered candidate sentence;

extract a first part-of-speech phrase and a second part-of-speech phrase from a syntax tree of each of the filtered candidate sentences to constitute a phrase set, calculate, based on an importance of each of the filtered candidate sentences, an importance of the first part-of-speech phrase and the second part-of-speech phrase that are extracted from the candidate sentences; and

select, from the phrase set and based on the importance of the first part-of-speech phrase and the second part-of-speech phrase that correspond to the candidate sentence, the first part-of-speech phrase and the second part-of-speech phrase that meet the preset condition as the summary phrase set.

20. The computer program product of claim 19 , wherein the computer-executable instructions further cause the apparatus to:

filter out noise in each of the candidate sentences to obtain a candidate word set corresponding to each of the candidate sentences, wherein each of the candidate sentences comprises a plurality of words, wherein a plurality of importance corresponds to the words; and

filter out, based on an importance of each of the words, a word whose importance is less than a preset threshold in the candidate word set to obtain each of the filtered candidate sentences.