Multiple-document summarization using document clustering
View Patent ↗Systems and methods are disclosed for summarizing multiple documents by generating a model of the documents as a mixture of document clusters, each document in turn having a mixture of sentences, wherein the model simultaneously representing summarization information and document cluster structure; and determining a loss function for evaluating the model and optimizing the model.
1. A method for summarizing multiple documents, comprising:
a. generating a model BUV T of the documents as a mixture of document clusters, each document in turn having a mixture of sentences, wherein U is a corresponding sentence-topic matrix, V is a corresponding document-topic matrix, and B is a corresponding term-sentence matrix;
b. simultaneously representing summarization information and document cluster structure in the model BUV T ;
c. determining a loss function l, where the loss function comprises
l ( U,V )= KL ( A∥BUV T )− In Pr ( U,V ),
where KL is the Kullback-Leibler divergence, and where A is a corresponding term-document matrix;
d. evaluating and optimizing the model by translating a summarization and clustering problem into minimizing the loss l between the given documents and model reconstructed terms using the maximum likelihood estimation task comprising
U
,
V
=
arg
min
U
,
V
l
(
U
,
V
)
;
e. generating a summary of the documents at the same time as clustering documents into a given size of targeted summarization based on the model BUV T and the maximum likelihood estimation task.
2. The method of claim 1 , comprising receiving a document language model for each document.
3. The method of claim 2 , wherein the document language model comprises a unigram language model.
4. The method of claim 1 , comprising extracting sentence candidates from the document and receiving a sentence language model for each sentence candidate.
5. The method of claim 4 , wherein the sentence language model comprises a unigram language model.
6. The method of claim 1 , comprising determining model parameters U and V from a document language model and a sentence language model.
7. The method of claim 1 , comprising generating a matrix A of features of the documents or a matrix B of the features of the sentences.
8. The method of claim 1 , wherein each column of the model BUV T comprises features of a corresponding document generated by the model with parameters U and V.
9. The method of claim 1 , comprising formulating the model BUV T to approximate the document language model.
10. A method for summarizing multiple documents, comprising:
a. receiving a document language model for each document, wherein the document language models are used to generate a corresponding term-document matrix A;
b. extracting sentence candidates from the documents and receiving a sentence language model for each sentence candidate, wherein the sentence language models are used to generate a corresponding term-sentence matrix B;
c. determining model parameters U and V from the document language models and the sentence language models, wherein U is a corresponding sentence-topic matrix, and V is a corresponding document-topic matrix;
d. simultaneously representing summarization information and document cluster structure in a model BUV T of the documents as a mixture of document clusters, each document in turn having a mixture of sentences;
e. minimizing loss according to a loss function l between the documents and model reconstructed terms according to a maximum likelihood estimation task comprising
U
,
V
=
arg
min
l
(
U
,
V
)
U
,
V
;
and
f. summarizing the documents at the same time as clustering documents into a given size of targeted summarization based on the model BUV T and the maximum likelihood estimation task.
11. The method of claim 10 , wherein the document or sentence language model comprises a unigram language model.
12. The method of claim 11 , wherein A is a corresponding matrix of features of the documents.
13. The method of claim 11 , wherein B is a corresponding matrix of features of the sentences.
14. The method of claim 10 , wherein each column BUV T of the model comprises features of a corresponding document generated by the model with parameters U and V.
15. The method of claim 10 , wherein the loss function l comprises a Kullback-Leibler divergence function or a Frobenius matrix norm.