KR-20260062817-A - AI solution for generating responses based on conversational characteristics extracted from past dialogues

Abstract

A conversation response generating device that generates a conversation response based on conversational characteristics and relationships using a language model according to the present invention may include: a memory storing a language model that generates response data representing a response that the user can transmit to the conversational speaker by reflecting the conversational characteristics of a past conversation performed between the conversational speaker conversing with the user and the user; and a processor that inputs real-time conversation data including dialogue sentences of a real-time conversation currently in progress between the conversational speaker and the user as first input data to the language model, and receives the response data as first output data from the language model.

Inventors

• 임요환

Assignees

• 임요환

Dates

Publication Date

20260507

Application Date

20250812

Claims (1)

1. A memory storing a language model that generates response data representing a response that the user can convey to the interlocutor, reflecting the conversational characteristics of a past conversation performed between the interlocutor conversing with the user and said user; and A processor that inputs real-time conversation data containing dialogue of a real-time conversation in progress between the conversationalist and the user as first input data to the language model, and receives the response data as first output data from the language model; The above language model A conversation feature extraction model that generates conversation feature data by extracting conversation features of the past conversation from past conversation data containing dialogue sentences of the past conversation; and A response generation model that generates the response data from the real-time conversation data; including The above conversation feature extraction model As the above conversation characteristic data, one or more of the following are generated: relationship characteristic data representing the relationship between the conversationalist and the user; conversationalist characteristic data representing characteristics by conversational sentence termination type for the conversationalist's conversational sentences; user characteristic data representing characteristics by conversational sentence termination type for the user's conversational sentences; emotion characteristic data representing the user's emotions in the past conversation; and context characteristic data representing the context of the past conversation. The above processor The language model is trained using the conversation characteristic data as training data so that the above response generation model generates the above response data by reflecting the characteristics of the above past conversation, and The above processor Calculate the number of response data selected by the user via the user's user dialogue for each response termination type, calculate a selection type ratio representing the ratio of each number of response termination types to the number of times the response data was selected by the user via the user's user dialogue, and identify the response termination type with the highest selection type ratio. The above processor After confirming the response termination type with the highest selection type ratio, if multiple response data are generated for the conversational text of the conversationalist, the display is controlled so that the response data of the response termination type with the highest selection type ratio is highlighted and displayed more than other response data. The relationship between the above conversational partner and the above user Characterized as being any one of the following: a relationship between a teacher and a parent, a romantic relationship, a relationship between a subordinate and a superior, a friendship, and a parent-child relationship. An AI solution that generates responses based on conversational characteristics extracted from past conversations.

Description

AI solution for generating responses based on conversational characteristics extracted from past dialogues The present invention relates to a conversation response generation device that generates conversation responses based on conversational characteristics and relationships using a language model. More specifically, the invention relates to a conversation response generation device capable of rapidly and appropriately generating responses reflecting the relationship between the conversationalist and the user, conversational content, emotions, and conversational habits, by training a language model using conversational characteristic data, which are conversational characteristics extracted from past conversations performed between the conversationalist and the user, as training data, and by having the language model generate response data representing responses that the user can convey to the conversationalist during a real-time conversation between the conversationalist and the user. Large AI, such as LLM (Large Language Model), is an artificial intelligence model trained using a very large amount of text data; it performs natural language processing tasks and can be used for various language modeling tasks. Most LLMs can be trained on datasets consisting of tens of billions of sentences containing various web documents and text data such as the internet, books, newspaper articles, and blogs, and can be used in various applications such as natural language understanding, sentence generation, machine translation, chatbots, and automatic summarization. Furthermore, deep learning models based on LLM are artificial intelligence technologies that learn human language and perform various tasks using that language. For example, a typical example is understanding sentences written by humans through natural language understanding and conducting conversations based on this. Conventional conversational systems store expected questions and answers in a large-capacity database and search the database for questions that match user input sentences to present answers. However, since conventional methods do not perform language processing, they recognize 'Please give me a television' and 'Please give me a TV' as different strings. Therefore, even if the question 'Please give me a television' and its answer exist in the database, there is a problem in that they cannot present an answer if the user says 'Please give me a TV'. Furthermore, conventional conversation systems have a problem in that they cannot handle goal-oriented conversations because they only process one-question, one-answer type dialogues that cannot maintain context. Conversations in person-to-person relationships follow a flow based on the topic, and words, predicates, and sentence-ending particles that constitute sentences are modified in various ways depending on the interlocutor. However, since existing conversation systems do not adaptively generate responses based on the conversation partner, modifications to particles or words do not occur, resulting in the generation of unnatural sentences as responses. Registered Patent No. 10-1359718 discloses a conversation management system and method. Registered Patent No. 10-1359718 relates to a method for constructing a model to effectively perform conversation management in building a chatting conversation system capable of free-form conversation with an artificial intelligence agent, rather than for a specific purpose. Registered Patent No. 10-1359718 provides a conversation service based on utterance pairs pre-stored within the system and does not consider conversational characteristics such as the relationship between the user and the conversational partner or speech style. Patent Application No. 10-2015-0086534 discloses a device and method for managing the conversational order according to conversational situations and topics. Patent Application No. 10-2015-0086534 relates to a method for determining a conversational topic by extracting keywords from a user's utterance information, considering the state of floor actions based on explicit signals including the user's gaze, gestures, and touch, and then applying a different conversational order according to a stored conversational model based on the conversational situation. Patent Application No. 10-2015-0086534 does not consider conversational characteristics such as the relationship between the user and the conversational partner or speech style. Patent Application No. 10-2013-0124534 discloses an interactive service device and method based on a user's speech style. Patent Application No. 10-2013-0124534 relates to an interactive service device and method based on a user's speech style, specifically a method in which the interactive service device analyzes the meaning of a user's spoken sentence to identify the user's speech intent and generates a response sentence based on the user's speech style. Patent Application No. 10-2013-0124534 does not cons