Human-machine interaction methods with validation

Abstract

The invention relates to a method for human-machine interaction, wherein a machine system (100) comprising a Large Language Model, LLM (104), reads in a text-based input (10), in particular a question, and the LLM (104) generates an answer (20) relating to the text-based input (10), wherein the answer (20) of the LLM (104) is validated by a self-reflection component (110) of the system (100).

Description

[0001] The invention relates to a method for human-machine interaction, wherein a machine system comprising a Large Language Model (LLM) reads in a text-based input, in particular a question, and the LLM generates an answer relating to the text-based input.

[0002] The invention can be used in text-based dialogue systems known as chatbots and in machine systems that have such a human-machine interface.

[0003] Large Language Models (LLMs), also known as large language models, are artificial intelligence systems designed to understand and generate text in human language. These LLMs can read text-based input from a human user, analyze it, and generate a response related to that input.

[0004] LLMs do not generate their responses based on the meaning or truth value of the respective statement, but rather on calculated probabilities for the words contained in the answer. When working with machine systems that exhibit such LLMs, the problem arises that the user cannot readily assess the truth value of the answer provided by the machine system.

[0005] Against this background, the task arises to expand the possibilities for the user to assess the quality of the LLM's response.

[0006] To solve the problem, a method for human-machine interaction according to claim 1 is proposed, wherein a machine system comprising a Large Language Model (LLM) reads in a text-based input, in particular a question, and the LLM generates an answer relating to the text-based input, wherein the answer of the LLM is validated by a self-reflection component of the system.

[0007] Validation through the system's self-reflection component allows for the quantification of the quality of the generated response. The user of the machine learning system can then better assess the quality of the response based on the validation results.

[0008] The self-reflection component can be configured as a self-learning component. It can observe multiple responses from the LLM (Learning Management Model). Initial values ​​can be provided to the self-reflection component. Based on these initial values, the self-reflection component can generate patterns, particularly through experimentation. Based on these generated patterns, the self-reflection component can create a model. If the self-reflection component observes and, if necessary, validates new responses from the LLM, it can modify previously generated patterns. Subsequently, the self-reflection component can generate a new model based on these modified patterns.

[0009] According to an advantageous embodiment of the invention, the validation process includes generating validation information expressed as a probability, which indicates the likelihood that the LLM's response is correct. Based on this generated probability, the user of the machine system can better assess whether or not they can trust the LLM's response.

[0010] According to an advantageous embodiment of the invention, the validation process includes generating validation information in the form of a domain identifier, which indicates the knowledge domain to which the LLM's response belongs. This domain identifier, particularly when considered in conjunction with the question, makes it easier for the user of the machine system to recognize whether the LLM has correctly analyzed the question. The user can compare the domain identifier with their existing knowledge of the question and the response. Furthermore, the user can interpret the response in light of the domain identifier.

[0011] According to an advantageous embodiment of the invention, the validation process includes generating validation information in the form of a location indicator, which specifies the location or environment to which the LLM's response is assigned. This location indicator can facilitate the user's interpretation of the LLM's response. Furthermore, the user can use the location indicator to verify whether the LLM has correctly analyzed the question.

[0012] According to an advantageous embodiment of the invention, the validation includes generating validation information in the form of a method specification, which indicates how the answer is obtained.

[0013] According to an advantageous embodiment of the invention, the validation process includes generating validation information in the form of a frequency indicator, which shows how often the LLM's response occurs. This frequency indicator can provide the user with information as to whether the statement contained in the LLM's response relates to a situation that occurs rather rarely or rather frequently. This information facilitates the user's assessment of the quality of the LLM's response. Preferably, the validation process includes generating a test result indicator that shows whether the generated frequency indicator lies within a predefined range, for example, between a predefined lower threshold and a predefined upper threshold.

[0014] According to an advantageous embodiment of the invention, the LLM response is output together with one or more validation pieces of information. By outputting both the response and the validation information, for example, including a probability of occurrence and / or a domain specification and / or a location specification and / or a method specification and / or a frequency specification, the user can be more easily comprehensible about the response and its quality.

[0015] According to an advantageous embodiment of the invention, the machine system comprises a database, and the response is generated based on the data contained in the database. The database can contain data that is directly accessed by the LLM (Launch Machine Learning Module), for example, by having the text-based input contain a question relating to data stored in the database, and the LLM's response comprising data read from the database. Alternatively or additionally, the database can contain data that is accessed to generate a data-dependent response from the LLM. For example, the text-based input can contain a question relating to data stored in the database, and the response can be generated based on the data contained in the database.

[0016] According to an advantageous embodiment of the invention, the database comprises operating data from at least one unit to be monitored, in particular a machine, and the response is generated based on the operating data contained in the database. Such a machine-based system enables condition monitoring, particularly with root cause analysis. The database can contain operating data from several units to be monitored, in particular several machines. This allows the machine-based system to generate responses relating to one or more units, in particular machines within a machine park, for example, a production facility. The operating data can be measured values ​​of physical quantities, such as temperature, position, vibration frequency, and / or vibration amplitude.Alternatively or additionally, the operating data can include status information, such as a machine state, which was determined, for example, by evaluating one or more measured values.

[0017] According to an alternative advantageous embodiment of the invention, the machine system is a chatbot.

[0018] The machine system can include an electronic data processing device, for example, a programmable computer. The LLM and the self-reflection component are preferably designed as software components that can be executed by the electronic data processing device.

[0019] Further details and advantages of the invention will be explained below with reference to the exemplary embodiment shown in the drawings. This shows: Fig. 1 A machine system in a block diagram to illustrate an embodiment of a human-machine interaction method according to the invention.

[0020] In the Fig. Figure 1 shows a block diagram of a machine system 100, which is designed as a system for monitoring the state of several machines A, B, C, D.

[0021] The machine system 100 includes a database 101 in which operating data of the machines A, B, C, D to be monitored are stored. The operating data can include measured values ​​from sensors arranged on the machines A, B, C, D, for example, measured values ​​of temperature and / or measured values ​​of the position of a machine element and / or measured values ​​of the vibration of a machine element, e.g., measured values ​​of a vibration frequency and / or a vibration amplitude.

[0022] Another component of the machine system 100 is a state model 105, which is created and / or updated based on the operational data contained in the database 101. The state model 105 can contain snapshots of different operating states of machines A, B, C, and D.

[0023] The machine system 100 also includes a Large Language Model (LLM) 104, which enables text-based communication with the system 100. A user of the system 100 can transmit a text-based input 10, in particular a question, to the system 100. The system 100 reads the question, and the LLM 104 then generates a text-based response 20 to the text-based input 10. The input 10 could, for example, be a question about the current state of a specific machine B. The LLM 104 generates a corresponding response 20, which informs the user of the state of machine B. The response 20 can also contain information about why the machine B is in the specified state and / or information about how the machine B can be changed from the specified state to another state.

[0024] The machine system 100 reads the text-based input 10 in conjunction with an action model repository 102, which contains a list of inputs 10 or user commands, each linked to an action to be performed. The LLM 107 reads the text-based input 10, analyzes it, and determines the intention behind it. Using a knowledge base 103 and a state model 105, an inference engine 104 derives a statement that depends on the text-based input 10.

[0025] Based on the statement derived by the inference engine 104, the LLM 107 generates the text-based response 20. This is displayed to the user.

[0026] According to the invention, the response 20 is fed to a self-reflection component of the machine system 110. The self-reflection component 110 validates the response 20 and generates validation information 30, which is output along with the response 20 and displayed to the user. The user of the machine system 110 can better assess the quality of the response 20 based on the validation information 30. The validation information can include a probability and / or a location and / or a method and / or a frequency and / or a domain. Reference symbol list: 10 text-based input 20 answers 30 Validation information 100 machine system 101 Database 102 Action Model Repository 103 Knowledge Base 104 Inference machine 105 State model 106 Action trigger 107 Large Language Model, LLM 110 Self-reflection component

Images