Adjustment of communication content

Abstract

A method for personalizing messages between a sender and a receiver is provided. The method includes semantically analyzing a communication history to form a knowledge graph, deriving a formality level value using a first trained ML model, analyzing parameter values ​​of replies to determine a receiver impact score, and training a second ML system to generate a model for predicting the receiver impact score value. The method also includes selecting a linguistic expression for a message to be drafted, determining an expressive intent, generating a modified linguistic expression by modifying the linguistic expression based on the formality level and the expressive intent, and testing whether the modified linguistic expression has an increased likelihood of having a higher receiver impact score. The method also includes repeating the steps of selecting a linguistic expression, determining the expressive intent, modifying the linguistic expression, and testing until a stopping criterion is met.

Claims

1. A method for personalizing a message between a sender and a receiver, wherein the method is The following steps are performed: semantically analyze at least the messages in the communication history, including the messages sent between the sender and the receiver; and, based on the results of the semantic analysis, form a knowledge graph representing the relationship between a sender identifier that identifies the sender and a receiver identifier that identifies the receiver. Using a first trained machine learning model that has been trained to derive a formality level value representing the relationship between two parties from a knowledge graph, the formality level value between the sender and the receiver is derived from the knowledge graph, Based on the parameter value of the reply in the communication history, which is a parameter value related to the recipient impact score value representing the impact of the received message, the recipient impact score value is determined such that the recipient impact score value is higher when the reply parameter value indicates that the recipient had a significant reaction to the message. The second machine learning system is trained to generate a model for predicting the recipient impact score value based on the knowledge graph and the formality level value, Select the linguistic expressions to be corrected from the drafted message, To determine the expressive intent of the selected linguistic expression, Based on the aforementioned formality level value and the intended expression, a modified language expression is generated by modifying the language expression to one that corresponds to the formality level value, without changing the intended expression in the language expression before and after modification. Using a trained third machine learning model that outputs prediction results indicating an increase or decrease in impact score values, the method determines, based on the prediction results, whether the modified language expression is more likely to result in a higher recipient impact score value, wherein the third machine learning model takes the language expression before modification, the modified language expression, and the output of the trained second machine learning system as inputs. The process of selecting the linguistic expression, determining the intent of the expression, modifying the linguistic expression, and determining a higher recipient impact score is repeated until the termination criteria are met. Methods that include...

2. The method according to claim 1, wherein the third machine learning model is a reinforcement learning model.

3. The method according to claim 1 or 2, wherein the training of the third machine learning model also includes using a bidirectional transformer to predict the expression intent.

4. The method according to claim 3, further comprising using the communication history, the knowledge graph, the formality level value, and the receiver impact score value as training data for the training of the bidirectional transformer.

5. In a one-to-many communication that transmits to multiple recipients, the recipient identifier is a plurality of recipient identifiers used to identify multiple users, The method according to any one of claims 1 to 4, wherein the modified language expression is generated based on the formality level values ​​between all of the plurality of recipients and the sender.

6. The modification of the linguistic expression is performed by at least one selected from the group consisting of word substitution, sentence structure reversal, content rearrangement, word deletion, message block reordering, use of synonyms, linguistic and stylistic adjustments used by the recipient in past communications, and GPT2 transformer modification. The method according to any one of claims 1 to 5, wherein the GPT2 transformer is trained to construct partial sentences in a personalized manner to generate paragraph segments in a personalized manner by synthesizing personalized text for the user.

7. The semantic analysis of the aforementioned communication history is The method according to any one of claims 1 to 6, comprising identifying a topic by a Latent Dirichlet Allocation (LDA) topic model.

8. The method according to any one of claims 1 to 7, wherein the formality level value is obtained by a bag-of-word model and the first trained machine learning model is a Gaussian naive Bayes classifier.

9. The method according to any one of claims 1 to 8, wherein the message is a text message or an audio message.

10. If there is no communication history between the sender and the receiver, The method according to any one of claims 1 to 9, wherein the communication history is replaced with a secondary communication history which is a communication history of communications between the recipient and a person other than the sender.

11. A message personalization system for personalizing messages between a sender and a receiver, wherein the message personalization system is Memory and The system includes a processor that communicates with the memory, and the processor The communication history, including messages transmitted between the sender and the receiver, is semantically analyzed, and based on the results of the semantic analysis, a knowledge graph is formed that represents the relationship between a sender identifier that identifies the sender and a receiver identifier that identifies the receiver. Using a first trained machine learning model that has been trained to derive a formality level value representing the relationship between two parties from a knowledge graph, the formality level value between the sender and the receiver is derived from the knowledge graph, Based on the parameter value of the reply in the communication history, which is a parameter value related to the recipient impact score value representing the impact of the received message, the recipient impact score value is determined such that the recipient impact score value is higher when the reply parameter value indicates that the recipient had a significant reaction to the message. The second machine learning system is trained to generate a model for predicting the recipient impact score value based on the knowledge graph and the formality level value, Select the linguistic expressions to be corrected from the drafted message, Determining the expressive intent of the aforementioned linguistic expression, Based on the aforementioned formality level value and the intended expression, a modified language expression is generated by modifying the language expression to one that corresponds to the formality level value, without changing the intended expression in the language expression before and after modification. Using a trained third machine learning model that outputs prediction results indicating an increase or decrease in impact score values, it is determined, based on the prediction results, whether the modified language expression is more likely to have a higher recipient impact score value, wherein the third machine learning model takes the language expression before modification, the modified language expression, and the output of the trained second machine learning system as input. The process of selecting the linguistic expression, determining the intent of the expression, modifying the linguistic expression, and determining a higher recipient impact score is repeated until the termination criteria are met. A message personalization system configured to perform actions including the following.

12. The message personalization system according to claim 11, wherein the third machine learning model is a reinforcement learning model.

13. The training of the third machine learning model described above is A message personalization system according to claim 11 or 12, comprising using a bidirectional transformer to predict the intended expression.

14. The training of the aforementioned bidirectional transformer is The message personalization system according to claim 13, comprising using the communication history, the knowledge graph, the formality level value, and the recipient impact score value as training data.

15. In a one-to-many communication that transmits to multiple recipients, the recipient identifier is a plurality of recipient identifiers used to identify multiple users, The message personalization system according to any one of claims 11 to 14, wherein the modified language expression is generated based on the formality level value and the expression intent between the sender and the plurality of users.

16. The modification of the linguistic expression is performed by at least one selected from the group consisting of word substitution, sentence structure reversal, content rearrangement, word deletion, message block reordering, use of synonyms, linguistic and stylistic adjustments used by the recipient in past communications, and GPT2 transformer modification. The message personalization system according to any one of claims 11 to 15, wherein the GPT2 transformer is trained to construct partial sentences in a personalized manner by synthesizing personalized text for the user to generate a paragraph segment.

17. The message personalization system according to any one of claims 11 to 16, wherein the semantic analysis of the communication history includes identifying topics by a Latent Dirichlet Allocation (LDA) topic model.

18. The message personalization system according to any one of claims 11 to 17, wherein the formality level value is obtained by a bag-of-word model and the first trained machine learning model is a Gaussian naive Bayes classifier.

19. A computer program for personalizing messages between a sender and a receiver, wherein the computer program includes program instructions that are executed by one or more computer systems. The process involves performing a semantic analysis on at least the messages in the communication history, including the messages sent between the sender and the receiver, and generating an analysis in which, based on the results of the semantic analysis, the analysis is adapted to form a knowledge graph representing the relationship between a sender identifier that identifies the sender and a receiver identifier that identifies the receiver. Using a first trained machine learning model that has been trained to derive a formality level value representing the relationship between two parties from a knowledge graph, the formality level value between the sender and the receiver is derived from the knowledge graph, Based on the parameter value of the reply in the communication history, which is a parameter value related to the recipient impact score value representing the impact of the received message, the recipient impact score value is determined such that the recipient impact score value is higher when the reply parameter value indicates that the recipient had a significant reaction to the message. The second machine learning system is trained to generate a model for predicting the recipient impact score value based on the knowledge graph and the formality level value, Select the linguistic expressions to be corrected from the drafted message, Determining the expressive intent of the aforementioned linguistic expression, Based on the aforementioned formality level value and the intended expression, a modified language expression is generated by modifying the language expression to one that corresponds to the formality level value, without changing the intended expression in the language expression before and after modification. The method involves using a trained third machine learning model that outputs a prediction result indicating an increase or decrease in impact score values, and determining, based on the prediction result, whether the modified language expression is more likely to have a higher recipient impact score value, wherein the third machine learning model takes the language expression before modification, the modified language expression, and the output of the trained second machine learning system as inputs. The program instructions for selection, determination, generation, and judgment are repeated until the stop criteria are met. A computer program that performs a certain action.

Images