Method and device for stepwise purification of sentences with generative artificial intelligence
Provided is a method of generating stepwise purified sentences using a deep learning-based interactive model. The method includes receiving input sentences by a user's terminal, detecting malicious words in the input sentences, and generating a plurality of stagewise purified sentences using a plurality of types of purification including purification of the malicious words according to stepwise sentence purification, by a language purification model part of a stepwise purified sentence generation system, and presenting, by the terminal, the plurality of stepwise purified sentences to a user to allow the user to select one of the plurality of stepwise purified sentences, and transmitting, by the terminal, the selected stepwise purified sentence to the stepwise purified sentence generation system to determine a final purified sentence.
1 . A method of generating stepwise purified sentences using a deep learning-based interactive model, the method comprising:
receiving input sentences by a user's terminal;
detecting malicious words in the input sentences, and generating a plurality of stepwise purified sentences using a plurality of types of purification including purification of the malicious words according to stepwise sentence purification by a language purification model part of a stepwise purified sentence generation system; and
presenting, by the terminal, the plurality of stepwise purified sentences to a user to allow the user to select one of the plurality of stepwise purified sentences, and transmitting, by the terminal, the selected stepwise purified sentence to the stepwise purified sentence generation system to determine a final purified sentence, wherein
the language purification model part includes a model repeatedly performing different types of sentence purification given at each stage, and
the different types of sentence purification include masking malicious words, replacing masked words with purified terms, and softening the entire sentence.
2 . The method of claim 1 , wherein the language purification model includes a reinforcement learning from human feedback (RLHF)-based large language model (LLM).
3 . The method of claim 2 , wherein the stepwise sentence purification is performed at least three times, and results from each stage of the method of stepwise purification of sentences are stored in a database, wherein the purified sentences stored in the database are presented to and selected by the user.
4 . The method of claim 1 , wherein stages of the method of stepwise purification of sentences are performed at least three times, and results from each stage of the method of stepwise purification of sentences are stored in a database, wherein the purified sentences stored in the database are presented to and selected by the user.
5 . The method of claim 1 , wherein the language purification model part includes a single model, and when a prompt for each stage from a prompt engineering part is input to the single model, the single model repeatedly performs the different types of sentence purification given at each stage.
6 . The method of claim 5 , wherein, after the terminal receives the input sentences, the stepwise purified sentence generation system pre-processes the input sentences, and then the language purification model part purifies the input sentences.
7 . The method of claim 1 , wherein the language purification model part includes a plurality of models specialized for each stage, and each model performs the different types of sentence purification given to each stage.
8 . The method of claim 7 , wherein, after the terminal receives the input sentences, the stepwise purified sentence generation system pre-processes the input sentences, and then the language purification model part purifies the input sentences.
9 . A stepwise purified sentence generation system that purifies input sentences from a user's terminal, the system comprising:
a model part configured to detect malicious words in the input sentences and generate a plurality of stepwise purified sentences using a plurality of types of purification including purification of the malicious words;
an output post-process part configured to store purified sentences from the model part in a database; and
a processing unit configured to detect the purified sentences in the database through polling and to send the purified sentences, generated from the model part and stored by the output post-process part, to the user's terminal,
wherein the stepwise purified sentence generation system is configured to determine the purified sentence selected through the terminal as a final purified sentence,
wherein the model part includes a model repeatedly performing different types of sentence purification given at each stage, and
wherein the different types of sentence purification include masking malicious words, replacing masked words with purified terms, and softening the entire sentence.
10 . The system of claim 9 , wherein the model part includes a reinforcement learning from human feedback (RLHF)-based large language model (LLM).
11 . The system of claim 9 , wherein the model part performs stepwise purified sentence generation at least three times, and stores the purified sentences generated at each stage in a database, wherein the purified sentences stored in the database are transmitted to the user's terminal.
12 . The system of claim 9 , wherein the model part includes a single model, and a prompt for each stage from a prompt engineering part is input to the single model so that the single model repeatedly performs the different types of sentence purification given at each stage.
13 . The system of claim 9 , wherein the model part includes a plurality of models specialized for each stage, and each model is configured to perform the different types of sentence purification given in each stage.
14 . The system of claim 9 , wherein an input pre-process part is provided on an input side of the model part, and an output post-process part is provided on an output side of the model part.