Method and apparatus for summarizing document based on document retrieval

1 . A method for summarizing a document, by an apparatus for summarizing a document, the method comprising:

acquiring a user inquiry vector from a user retrieval query;

extracting a key sentence of a passage of a document stored in a retrieval database from the passage of the document;

calculating a first similarity score between the user inquiry vector and a sentence vector corresponding to the key sentence extracted from the passage of the document by inputting the user inquiry vector and the sentence vector to a bi-encoder of a bi-encoder type deep learning model;

calculating a second similarity score between the user inquiry vector and a sentence vector corresponding to a sentence summarizing the passage of the document stored in the retrieval database;

generating a question from the passage of the document stored in the retrieval database through a generation model;

calculating a third similarity score between a question vector corresponding to the question generated from the passage stored in the retrieval database and the user inquiry vector;

calculating a first weighted score based on the first similarity score, the second similarity score, and the third similarity score;

determining a first document candidate group including first documents based on the calculated first weighted score;

calculating a first score indicating similarity between the user retrieval query and a passage stored in the retrieval database through a phrase matching;

calculating a second score indicating similarity between key query information including a keyword of the user retrieval query extracted through a user query analysis module and a keyword included in the passage stored in the retrieval database;

calculating a third score indicating similarity between the user retrieval query and the passage stored in the retrieval database through a shingle matching;

calculating a second weighted score based on the first score, the second score, and the third score;

determining a second document candidate group including second documents based on the calculated second weighted score;

acquiring a primary document candidate group including the first documents of the first document candidate group and the second documents of the second document candidate group;

calculating a fourth score indicating similarity between a passage of a document in the primary document candidate group and the user retrieval query by inputting the passage of the document in the primary document candidate group and the user retrieval query to a cross-encoder of cross-encoder type deep learning model;

calculating a fifth score indicating similarity between the key query information including the keyword of the user retrieval query and a keyword included in the passage of the document in the primary document candidate group using a score calculation algorithm;

calculating a weighted sum similarity score based on the fourth score and the fifth score and adjusting a ranking of documents in the primary document candidate group based on the weighted sum similarity score;

determining a summarization target document based on the adjusted ranking of the documents; and

generating an abstractive summarization-based first summary through a first output layer of a trained document summarization model and an extractive summarization-based second summary through a second output layer of the trained document summarization model by inputting the determined summarization target document to the trained document summarization model.

2 . The method of claim 1 , wherein the acquiring the user inquiry vector from the user retrieval query includes:

acquiring the user inquiry vector in a unit of sentence from the user retrieval query.

3 . The method of claim 1 , further comprising:

generating a question to specify the user retrieval query based on the user retrieval query using the generation model when the user retrieval query is not recognized,

transmitting the generated question to a user terminal, and

receiving a user retrieval query corresponding to the transmitted question from the user terminal.

4 . The method of claim 1 , wherein the score calculation algorithm includes a BM25F score calculation algorithm.

5 . The method of claim 1 , wherein the determining the summarization target document based on the adjusted ranking includes:

determining a document in order of the adjusted ranking among the documents included in the primary document candidate group as the summarization target document.

6 . The method of claim 1 , wherein the generating of the abstractive summarization-based first summary and the extractive summarization-based second summary includes:

inputting the determined summarization target document to an input layer of the trained document summarization model;

acquiring a first summarization vector through the first output layer of the trained document summarization model;

generating the abstractive summarization-based first summary based on the first summarization vector;

acquiring a second summarization vector through the second output layer of the trained document summarization model; and

generating the extractive summarization-based second summary based on the second summarization vector.

7 . A non-transitory computer-readable recording medium in which a computer program executed by a computer to perform operations is recorded, the operations comprising:

acquiring a user inquiry vector from the user retrieval query;

extracting a key sentence of a passage of a document stored in a retrieval database from the passage of the document;

calculating a first similarity score between the user inquiry vector and a sentence vector corresponding to the key sentence extracted from the passage of the document by inputting the user inquiry vector and the sentence vector to a bi-encoder of a pre-trained bi-encoder type deep learning model;

calculating a second similarity score between the user inquiry vector and a sentence vector corresponding to a sentence summarizing the passage of the document stored in the retrieval database;

generating a question from the passage of the document stored in the retrieval database through a generation model;

calculating a third similarity score between a question vector corresponding to the question generated from the passage stored in the retrieval database and the user inquiry vector;

calculating a first weighted score based on the first similarity score, the second similarity score, and the third similarity score;

determining a first document candidate group including first documents based on the calculated first weighted score;

calculating a first score indicating similarity between the user retrieval query and a passage stored in the retrieval database through a phrase matching;

calculating a second score indicating similarity between key query information including a keyword of the user retrieval query extracted through a user query analysis module and a keyword included in the passage stored in the retrieval database;

calculating a third score indicating similarity between the user retrieval query and the passage stored in the retrieval database through a shingle matching;

calculating a second weighted score based on the first score, the second score, and the third score;

determining a second document candidate group including second documents based on the calculated second weighted score;

acquiring a primary document candidate group including the first documents of the first document candidate group and the second documents of the second document candidate group;

calculating a fourth score indicating similarity between a passage of a document in the primary document candidate group and the user retrieval query by inputting the passage of the document in the primary document candidate group and the user retrieval query to a cross-encoder of cross-encoder type deep learning model;

calculating a fifth score indicating similarity between the key query information including the keyword of the user retrieval query and a keyword included in the passage of the document in the primary document candidate group using a score calculation algorithm;

calculating a weighted sum similarity score based on the fourth score and the fifth score and adjusting a ranking of documents in the primary document candidate group based on the weighted sum similarity score;

determining a summarization target document based on the adjusted ranking of the documents; and

generating an abstractive summarization-based first summary through a first output layer of a trained document summarization model and an extractive summarization-based second summary through a second output layer of the trained document summarization model by inputting the determined summarization target document to the trained document summarization model.

8 . An apparatus for summizing a document, comprising:

a memory storing instructions; and

a hardware processor configured to execute the instructions to:

acquire a user inquiry vector from a user retrieval query,

extract a key sentence of a passage of a document stored in a retrieval database from the passage of the document,

calculate a first similarity score between the user inquiry vector and a sentence vector corresponding to the key sentence extracted from the passage of the document by inputting the user inquiry vector and the sentence vector to a bi-encoder of a bi-encoder type deep learning model,

calculate a second similarity score between the user inquiry vector and a sentence vector corresponding to a sentence summarizing the passage of the document stored in the retrieval database,

generate a question from the passage of the document stored in the retrieval database through a generation model,

calculate a third similarity score between a question vector corresponding to the question generated from the passage stored in the retrieval database and the user inquiry vector,

calculate a first weighted score based on the first similarity score, the second similarity score, and the third similarity score,

determine a first document candidate group including first documents based on the calculated first weighted score,

calculate a first score indicating similarity between the user retrieval query and a passage stored in the retrieval database through a phrase matching,

calculate a second score indicating similarity between key query information including a keyword of the user retrieval query extracted through a user query analysis module and a keyword included in the passage stored in the retrieval database,

calculate a third score indicating similarity between the user retrieval query and the passage stored in the retrieval database through a shingle matching,

calculate a second weighted score based on the first score, the second score, and the third score,

determine a second document candidate group including second documents based on the calculated second weighted score,

acquire a primary document candidate group including the first documents of the first document candidate group and the second documents of the second document candidate group,

calculate a fourth score indicating similarity between a passage of a document in the primary document candidate group and the user retrieval query by inputting the passage of the document in the primary document candidate group and the user retrieval query to a cross-encoder of cross-encoder type deep learning model,

calculate a fifth score indicating similarity between the key query information including the keyword of the user retrieval query and a keyword included in the passage of the document in the primary document candidate group using a score calculation algorithm,

calculate a weighted sum similarity score based on the fourth score and the fifth score and adjust a ranking of documents in the primary document candidate group based on the weighted sum similarity score,

determine a summarization target document based on the adjusted ranking of the documents, and

generate an abstractive summarization-based first summary through a first output layer of a trained document summarization model and an extractive summarization-based second summary through a second output layer of the trained document summarization model by inputting the determined summarization target document to the trained document summarization model.

9 . The apparatus of claim 8 , wherein the hardware processor is further configured to acquire the user inquiry vector in a unit of sentence from the user retrieval query.

10 . The apparatus of claim 8 , wherein the hardware processor is further configured to:

generate a question to specify the user retrieval query based on the user retrieval query using the generation model when the user retrieval query is not recognized,

transmit the generated question to a user terminal, and

receive a user retrieval query corresponding to the transmitted question from the user terminal.

11 . The apparatus of claim 8 , wherein the score calculation algorithm includes a BM25F score calculation algorithm.

12 . The apparatus of claim 8 , wherein the hardware processor is further configured to:

determine a document in order of the adjusted ranking among the documents included in the primary document candidate group as the summarization target document.

13 . The apparatus of claim 8 , wherein the hardware processor is further configured to:

input the determined summarization target document to an input layer of the trained document summarization model,

acquire a first summarization vector through the first output layer of the trained document summarization model,

generate the abstractive summarization-based first summary based on the first summarization vector,

acquire a second summarization vector through the second output layer of the trained document summarization model, and

generate the extractive summarization-based second summary based on the second summarization vector.