System and method for automatically expanding input text

Patent metadata

Field	Value
Publication number	US-10402494-B2
Application number	US-201715439416-A
Country	US
Kind code	B2
Filing date	Feb 22, 2017
Priority date	Dec 6, 2016
Publication date	Sep 3, 2019
Grant date	Sep 3, 2019

Abstract

Provided is a method of automatically expanding input text. The method includes receiving input text composed of a plurality of documents, extracting a sentence pair that is present in different documents among the plurality of documents, setting the extracted sentence pair as an input of an encoder of a sequence-to-sequence model, setting an output of the encoder as an output of a decoder of the sequence-to-sequence model and generating a sentence corresponding to the input, and generating expanded text based on the generated sentence.

First claim

What is claimed is:

1. A method of automatically expanding input text, the method comprising:
receiving input text composed of a plurality of documents;
extracting a sentence pair that is present in different documents among the plurality of documents;
setting the extracted sentence pair as an input of an encoder of a sequence-to-sequence model;
setting an output of the encoder as an output of a decoder of the sequence-to-sequence model and generating a sentence corresponding to the input; and
generating expanded text based on the generated sentence,
wherein the extracting of a sentence pair that is present in different documents among the plurality of documents comprises:
randomly extracting a plurality of sentence pairs that are present in different documents among the plurality of documents; and
sorting the plurality of sentence pairs based on a distance between words of two sentences constituting each of the plurality of sentence pairs.

2. The method of claim 1 , further comprising:
receiving learning text including a plurality of documents composed of a plurality of sentences;
extracting a sentence chain from the learning text; and
generating the sequence-to-sequence model, based on the extracted sentence chain, wherein
the sentence chain is composed of a set including three sentences related to one document.

3. The method of claim 2 , wherein the extracting of a sentence chain from the learning text comprises extracting the sentence chain from each of the plurality of documents included in the learning text in sequence.

4. The method of claim 3 , wherein the extracting of a sentence chain from the learning text comprises:
extracting a word chain from one sentence included in one of the plurality of documents;
storing the extracted word chain in a word chain candidate list;
determining whether the word chain is extracted from all words included in the one sentence;
extracting at least one sentence chain corresponding to the word chain included in the word chain candidate list and storing the extracted sentence chain in a sentence chain list when the word chain is extracted from all of the words;
determining whether the sentence chain is extracted from all sentences included in the one document; and
outputting a sentence chain list corresponding to the document when the sentence chain is extracted from all of the sentences.

5. The method of claim 4 , wherein:
the words included in the sentence are set as a vector value through word embedding; and
positions of the words on a vector space are determined based on similarity between contexts thereof.

6. The method of claim 5 , wherein the extracting of a word chain comprises:
detecting a word positioned closest to any one of the words included in the sentence on the vector space to generate a partial word chain; and
detecting a word positioned closest to the partial word chain on the vector space to extract the word chain.

7. The method of claim 4 , wherein the extracting and storing of at least one sentence chain in a sentence chain list comprises sorting the sentence chain to correspond to a word chain selected from among word chains included in the word chain candidate list based on a priority that is based on a predetermined determination criterion and storing the sorted sentence chain in the sentence chain list.

8. The method of claim 2 , wherein the generating of the sequence-to-sequence model comprises:
selecting a sentence chain for any one of the plurality of documents included in the learning text;
setting two of the three related sentences included in the sentence chain as the input of the encoder of the sequence-to-sequence model and setting the remaining sentence as the output of the decoder; and
learning the sentences set as the input and output to generate the sequence-to-sequence model.

9. The method of claim 1 , wherein the extracting of a sentence pair that is present in different documents among the plurality of documents comprises:
sorting the plurality of sentence pairs based on a distance between words of two sentences constituting each of the plurality of sentence pairs.

10. The method of claim 1 , wherein the setting of the extracted sentence pair as an input of an encoder of a sequence-to-sequence model comprises setting the sorted sentence pairs as the input of the encoder in sorting order.

11. The method of claim 10 , wherein the generating of expanded text based on the generated sentence comprises:
storing the generated sentence in a text expansion candidate list;
filtering the text expansion candidate list based on similarity to a pre-generated language model; and
generating the filtered text expansion candidate list as the expanded text.

12. The method of claim 11 , further comprising shuffling the plurality of sentences included in the input text, wherein
after the generated sentence is stored in the text expansion candidate list and then included in the plurality of sentences, the shuffling is performed on the plurality of sentences.

13. The method of claim 2 , wherein the extracting of a sentence pair that is present in different documents among the plurality of documents comprises:
embedding the extracted sentence pair based on a pre-learned recurrent neural network language model and expressing the embedded sentence pair as a vector; and
learning a 1-hop model configured to use a sentence chain including two sentences constituting the sentence pair and a counter sentence chain corresponding to the sentence chain to classify the sentence chain and the counter sentence chain, wherein
the 1-hop model configures the embedded sentence through layer P 1 and layer H 1 as a deep neural network model and obtains a resultant value to determine the sentence chain through layer O 1 .

14. The method of claim 13 , wherein the extracting of a sentence pair that is present in different documents among the plurality of documents comprises:
randomly extracting a plurality of sentence pairs that are present in different documents among the plurality of documents; and
sorting the plurality of sentence pairs based on the 1-hop model.

15. The method of claim 13 , further comprising learning a 2-hop model configured to use a sentence chain including the three related sentences and a counter sentence chain corresponding to the sentence chain to classify the sentence chain and the counter sentence chain, wherein
the 2-hop model has an output value of the layer H 1 of the 1-hop model and an embedded sentence that is not included in the sentence pair as an input of layer P 2 and obtains a resultant value to determine the sentence chain through layer H 2 and layer O 2 .

16. The method of claim 13 , wherein the recurrent neural network language model is any one of a Long Short-Term Memory (LSTM) and a Gated Recurrent Unit (GRU).

17. A text expansion system configured to automatically expand input text, the text expansion system comprising:
a communication module configured to transmit and receive data to and from an external device;
a memory configured to store a program for generating expanded text from the input text; and
a processor configured to execute the program stored in the memory, wherein,
by executing the program, the processor extracts a sentence pair that is present in different documents among a plurality of documents when input text composed of the plurality of documents is received, inputs the extracted sentence pair to an encoder of a sequence-to-sequence model, generates a sentence corresponding to the input as an output of a decoder of

Assignees

• Electronics & Telecommunications Res Inst

Inventors

• Chung Eui Sok
• Kang Byung Ok
• Park Ki Young
• Park Jeon Gue
• Song Hwa Jeon
• Lee Sung Joo
• Lee Yun Keun
• Jeon Hyung Bae

Classifications

• G06F40/56 Primary

  Natural language generation · CPC title

• G06F40/289 Primary

  Phrasal analysis, e.g. finite state techniques or chunking · CPC title

• G06F17/2881

  Physics · mapped topic

• G06F17/2775 Primary

  Physics · mapped topic

• G06N3/0442

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