KR-20260062512-A - METHOD AND DEVICE FOR GENERATING ARTIFICIAL INTELLIGENCE MODEL USING CONVERSATION CONTENT COMPRESSION, AND METHOD AND DEVICE FOR GENERATING ANSWER USING ARTIFICIAL INTELLIGENCE MODEL

KR20260062512AKR 20260062512 AKR20260062512 AKR 20260062512AKR-20260062512-A

Abstract

One embodiment of the present disclosure provides a method for generating an artificial intelligence model. The method comprises: a step of generating a training data set comprising a plurality of different conversation pairs, each consisting of a question and an answer matched to the question; a step of generating a second compressed data at a specific point in time based on a first compressed data of conversation content up to a point in time prior to a specific point in time and a conversation pair at the specific point in time; and a step of generating an expected answer to a question input at a point in time following the specific point in time using an artificial intelligence model trained to output an answer to a question input based on the training data set and the second compressed data.

Inventors

송현오
김장현

Assignees

서울대학교산학협력단

Dates

Publication Date: 20260507
Application Date: 20241029

Claims (16)

In a method for generating an artificial intelligence model performed by one or more processors, a) generating a training data set comprising multiple different conversation pairs consisting of a question and an answer matched to the question; b) generating a second compressed data at a specific point in time based on a first compressed data of conversation content up to a point in time prior to a specific point in time and a pair of conversations at the specific point in time; and c) A method for generating an artificial intelligence model, comprising the step of generating an expected answer to a question input at the next time point after the specific time point, using an artificial intelligence model trained to output an answer to a question input based on the above training data set and the above second compressed data.
In paragraph 1, A method for generating an artificial intelligence model, further comprising the step of retraining the artificial intelligence model by comparing the above-mentioned expected answer with the above-mentioned actual answer to the above-mentioned question so that the above-mentioned expected answer has a similarity of at least a certain level with the above-mentioned actual answer.
In paragraph 1, The above first compressed data and the above second compressed data are, An artificial intelligence model generation method that is generated based on forward operations with a Transformer language model.
In paragraph 1, The above step b) is, The method includes the step of generating the second compressed data by adding a compression token to each of the question and answer included in the conversation pair above, A method for generating an artificial intelligence model, wherein the second compressed data is based on keywords extracted from each of the questions and answers included in the conversation pair.
In paragraph 1, The above step b) is, A method for generating an artificial intelligence model, further comprising the step of storing the second compressed data.
Communication module; At least one processor; and It includes a memory electrically connected to the processor and storing at least one code executed in the processor, When the above memory is executed through the above processor, the processor, An artificial intelligence model generation device that generates a learning data set comprising multiple different conversation pairs consisting of a question and an answer matched to the question, generates a second compressed data at a specific point in time based on a first compressed data of conversation content up to a point in time prior to a specific point in time and a conversation pair at the specific point in time, and stores code that causes to generate an expected answer to a question input at the next point in time following the specific point in time by using an artificial intelligence model trained to output an answer to a question input based on the learning data set and the second compressed data.
In paragraph 6, The above memory allows the processor, An artificial intelligence model generation device further comprising the step of retraining the artificial intelligence model by comparing the above-mentioned expected answer with the above-mentioned actual answer to the above-mentioned question so that the above-mentioned expected answer has a similarity of at least a certain level with the above-mentioned actual answer.
In paragraph 6, The above first compressed data and the above second compressed data are, An artificial intelligence model generation device that is generated based on forward operation with a Transformer language model.
In a method for generating answers using an artificial intelligence model executed by a server, i) a step of receiving a question at a specific point in time; and ii) includes the step of generating an answer to the above question using an artificial intelligence model, and The above artificial intelligence model is, An artificial intelligence model trained to output an answer to an input question based on a learning dataset comprising multiple different conversation pairs consisting of a question and an answer matched to the said question, a first compressed data regarding conversation content up to a point in time prior to a specific point in time, and a second compressed data of a specific point in time generated based on the conversation pairs of said specific point in time, wherein The above artificial intelligence model is, A method for generating an answer using an artificial intelligence model, wherein the artificial intelligence model is retrained by comparing the answer output based on the artificial intelligence model with the actual answer to the question so that the answer has a certain degree of similarity or greater than the actual answer.
In Paragraph 9, iii) A method for generating an answer using an artificial intelligence model, further comprising the step of transmitting the above answer to a terminal connected to the server.
In Paragraph 9, The above first compressed data and the above second compressed data are, A method for generating answers using an artificial intelligence model, which is generated based on forward operations with a Transformer language model.
In Paragraph 9, The above artificial intelligence model is, A method for generating answers using an artificial intelligence model, wherein a compression token is added to each of the questions and answers included in the above conversation pair to generate the above second compressed data, and the above second compressed data is based on keywords extracted from each of the questions and answers included in the above conversation pair.
In Paragraph 9, A method for generating an answer using an artificial intelligence model, further comprising the step of compressing the answer generated using the above question and the above artificial intelligence model to generate new compressed data, and updating the above artificial intelligence model based on the generated new compressed data.
Communication module; At least one processor; and It includes a memory electrically connected to the processor and storing at least one code executed in the processor, When the above memory is executed through the above processor, the processor, Store code that receives a question at a specific point in time and causes an artificial intelligence model to generate an answer to the said question, An answer generation device using an artificial intelligence model, wherein the artificial intelligence model is trained to output an answer to an input question based on a learning data set comprising multiple different conversation pairs consisting of a question and an answer matched to the question, a first compressed data regarding conversation content up to a point in time prior to a specific point in time, and a second compressed data of a specific point in time generated based on a conversation pair at the specific point in time, and wherein the artificial intelligence model is retrained so that the answer output based on the artificial intelligence model has a similarity of at least a certain degree to the actual answer to the question by comparing the answer output based on the artificial intelligence model with the actual answer to the question.
In Paragraph 14, The above memory allows the processor, An answer generation device using an artificial intelligence model, which compresses the answer generated using the above question and the above artificial intelligence model to generate new compressed data, and stores code that causes the artificial intelligence model to be updated based on the generated new compressed data.
In Paragraph 14, The above memory allows the processor, An answer generation device using an artificial intelligence model, which stores code that causes the above answer to be transmitted to a terminal connected to the communication module.

Description

Method and device for generating an artificial intelligence model using conversation content compression, and, method and device for generating an answer using an artificial intelligence model The present invention relates to a method and apparatus for generating an artificial intelligence model using conversation content compression, and a method and apparatus for generating an answer using an artificial intelligence model. More specifically, the invention relates to a method and apparatus for generating compressed data based on different conversation pairs consisting of answers matched to a question, and generating an answer to a question using an artificial intelligence model trained based on the compressed data and the conversation pairs. Existing technologies use a method of compressing context into text form, which results in poor compression performance and causes additional computational costs for language models to process the compressed text. Additionally, existing technologies have the disadvantage that compression costs increase as text accumulates because they must repeatedly compress all past text. For example, some studies on existing text-based compression methods have attempted to reduce required memory and computational load by compressing text token inputs into token form. However, this method performs compression within the space of the input tokens, resulting in low compression efficiency and inefficiency due to the need to process the compressed tokens again. As another example, an attempt was made to reduce inference costs by generating a compressed phenotype of a fixed input prompt using an existing fixed phenotype-based compression method. While this method is effective for fixed text that is used repeatedly, it is difficult to apply in cases where text accumulates due to continuous user interaction, such as in ChatGPT. This approach has the problem that all previous text must be compressed again, and consequently, the amount of computation required for compression continuously increases as interaction progresses. Accordingly, there is a need for technology capable of generating compressed phenotypes in real-time only for continuously provided text. FIG. 1 is a drawing illustrating a server and a terminal connected to it in communication according to an embodiment of the present invention. Figure 2 is a diagram illustrating the detailed configuration of the server shown in Figure 1. FIG. 3 is a diagram illustrating an example of an implementation of a server according to one embodiment of the present invention. FIG. 4 is a diagram illustrating the processing performance of a server according to one embodiment of the present invention. FIG. 5 is a diagram illustrating an example of application of a server according to one embodiment of the present invention. FIG. 6 is a flowchart illustrating the sequence of an artificial intelligence model generation method according to another embodiment of the present invention. FIG. 7 is a flowchart illustrating the sequence of a method for generating answers using an artificial intelligence model according to another embodiment of the present invention. The present disclosure will be described in detail below with reference to the attached drawings. However, the present disclosure may be implemented in various different forms and is not limited to the embodiments described herein. Furthermore, the attached drawings are intended only to facilitate understanding of the embodiments disclosed herein, and the technical concept disclosed herein is not limited by the attached drawings. All terms used herein, including technical and scientific terms, should be interpreted in the sense generally understood by those skilled in the art to which the present disclosure pertains. Terms defined in advance should be interpreted as having additional meanings consistent with relevant technical literature and the present disclosure, and should not be interpreted in a highly ideal or restrictive sense unless otherwise defined. In order to clearly explain the present disclosure in the drawings, parts unrelated to the explanation have been omitted, and the size, form, and shape of each component shown in the drawings may be varied. Throughout the specification, identical or similar parts are denoted by identical or similar reference numerals. In the following description, suffixes such as "module" and "part" for components are assigned or used interchangeably solely for the ease of drafting the specification, and do not inherently possess distinct meanings or roles. Furthermore, in describing the embodiments disclosed in this specification, detailed descriptions of related prior art have been omitted where it is determined that such detailed descriptions could obscure the essence of the embodiments disclosed in this specification. Throughout the specification, when it is stated that a part is "connected (connected, contacted, or coupled)" to another part, this includ