Controlling household appliance using language module

Abstract

A method of controlling a household appliance with a language module comprising a large language model (with a language module to control the household appliance), the method comprising:-the language module receiving an input query; -the language module determines whether the input query comprises a task-based query; -the language module translates the task-based query into a machine syntax, wherein the machine syntax is understandable by the home appliance; the language module outputs the machine grammar; -the household appliance receiving the machine syntax; and-the household appliance performs the machine syntax.

Description

Technical Field

[0001] The present invention relates to a method for controlling household appliances using a language module. The present invention further relates to a method for prompting and / or training a large language model to control household appliances using a language module. The present invention further relates to a method for generating a prompt and / or training dataset for a large language model. The present invention further relates to a prompt and / or training dataset for use in a method for training a large language model. The present invention further relates to the application of the method. The present invention further relates to a data processing system including a device for executing the method. The present invention further relates to a computer program. The present invention further relates to a computer-readable medium. Background Art

[0002] Today, it is possible to control household appliances using a voice control system, which includes a language module. According to the prior art, the language module acts on the transcription of a user command in a simple "intent and slot" paradigm. The intent can be, for example, "configure_washer_program" (configure the washing machine program), "start_program" (start the program), "stop_program" (stop the program), "get_remaining_program_time" (get the remaining program time), etc. The slot can be, for example, "program_name" (program name) as the slot name, and "cotton" (cotton fabric), "mix" (mixed fabric), "silk" (silk) etc. as the slot values, or "temperature" (temperature) as the slot name and "20", "30", "90" etc. as the slot values, or "spin_speed" (spin speed) as the slot name and "0", "800", "1000" etc. as the slot values, or "options" (options) as the slot name and one or more of "water_plus" (add water), "low_wrinkles" (low wrinkles) etc. as the slot values, and so on. Thus, the intent represents the use case or main intention of what the user is trying to achieve, particularly using a voice command, and the slot can further specify the intent.

[0003] Preferably, each command is associated with exactly one intention. Preferably, some commands, and especially some intentions, are associated with a slot, preferably with more than one slot. If a command concerns a subject outside the field of known intentions and / or slots, the language module will generate a special intention indicating that the command is not recognized. It is conceivable that such communication between the user and the household appliance is very restricted and is limited to predefined keywords, especially the selection and variation of predefined vocabulary (especially as in words or synonyms), and the selection and variation of predefined syntax (especially as in sentence structure). Summary of the Invention

[0004] The basic object of the present invention is to provide an improved, especially robust, technique for controlling household appliances using a language module. The present invention solves the above object by the subject matter of the independent claims. The dependent claims describe preferred embodiments.

[0005] A method, especially a computer-implemented method, for controlling a household appliance using a language module including a large language model, the method comprising the following steps: - The language module receives an input query, especially a natural language query; - The language module determines whether the input query includes a task-based query; - The language module translates the task-based query into a machine grammar, where the machine grammar is understandable, especially understandable by the household appliance; - The language module outputs the machine grammar; - The household appliance receives the machine grammar; and - The household appliance executes the machine grammar.

[0006] Preferably, the task-based query includes a query related to the user intention of changing the internal state of the household appliance. Preferably, the task-based query can be, but is not limited to, queries such as changing the selected program name or other program settings, such as for example "I want to wash a silk shirt with oil stains" or "Set the cotton fabric program to 60 degrees". Preferably, the language module outputs the corresponding machine grammar, especially setting the washing machine to the silk shirt washing program or the 60-degree cotton fabric program.

[0007] In particular, by using a language module including a large language model, a robust technique for controlling household appliances using the language module. Robustness can mean, for example, that commands, especially voice commands, especially natural language commands, can be understood by the language module regardless of which vocabulary (especially as in words and / or synonyms) and variations the command has been formulated with, and regardless of which syntax (especially as in sentence structures) the command has been formulated with. In particular, these commands can be understood by the household appliances by the language model translating task-based queries into machine grammar.

[0008] Preferably, what the language module outputs is that the machine grammar has been sent to the household appliance. In particular, this is the confirmation and / or feedback that the command has been understood by the language module. In particular, this is the signal that the household appliance has been instructed using the command.

[0009] Preferably, the output, especially the output from the language module regarding that the machine grammar has been sent to the household appliance, is converted into speech, especially voice output, especially converted into speech in a text-to-speech module. Preferably, this can be audible to the user, especially audibly in terms of acoustics.

[0010] Preferably, the language module determines whether the input query includes a secondary query. Preferably, the language module outputs an answer to the secondary query, especially a natural language answer. In particular, the language module answers commands for specific tasks related to household appliances and answers informative questions, especially secondary queries. Using such a language module, preferably, natural communication can be established.

[0011] Preferably, the secondary query includes queries related to auxiliary questions associated with any household appliance, and these auxiliary questions are preferably questions that the user may have and / or preferably have nothing to do with the user's intention to change the internal state of the household appliance, such as, for example, "How do I clean my washing machine?". Preferably, the language module answers using the corresponding output. However, preferably, the machine does not take any action, especially because it is not a task-based query but only a secondary query.

[0012] In particular, the input query may include a task-based query and an auxiliary query. In particular, the language module determines whether the input query includes a task-based query and / or an auxiliary query. Preferably, the language model outputs a machine grammar and / or an answer to the auxiliary query. For example, if the input query is "I want to wash a silk shirt with oil stains" or "Set the cotton fabric program to 60 degrees", the system response to the task-based query includes setting the washing machine to the silk shirt washing program or the 60-degree cotton fabric program, where the system response to the auxiliary query is a system answer such as "I have set the washing machine to the silk shirt washing program because this will clean the silk shirt and also help remove the oil stains" or "I have set the 60-degree cotton fabric program; please remind to only put cotton clothes in the washing machine".

[0013] Preferably, the household appliance is voice-controlled using voice commands. Preferably, the voice command is converted into an input query, particularly in an automatic speech recognition module. Preferably, the input query is in text form.

[0014] A method, particularly a computer-implemented method, for prompting and / or training a large language model to control a household appliance using a language module including the large language model, particularly for use with the method according to one of the preceding claims; the method includes: - Receiving an input prompt and / or a training data set, the input prompt and / or training data set including a connection between an input query, particularly an input query in natural language, and a machine grammar.

[0015] Preferably, a pre-trained large language model can be used as a basis. In the case of prompting and / or training a preferably pre-trained large language model, the large language model is capable of converting a command into a machine grammar, particularly a machine grammar understandable by the household appliance. Training can also be understood as adjusting and / or fine-tuning the pre-trained large language model.

[0016] Preferably, the input prompt and / or training data set is derived from documentation, particularly documentation related to the household appliance, which can be a manual of the household appliance software interface and / or internal and / or public documentation. For example, the internal documentation can be the documentation of the household appliance software interface that allows controlling the household appliance. In particular, existing manuals can be used to prompt and / or train a large language model, particularly an existing large language model. The software interface documentation preferably provides a link between natural language (particularly commands) and machine grammar. Preferably, the large language model not only recognizes the given machine grammar but also recognizes different commands that vary only in vocabulary and / or syntax and can be used as descriptions of the machine grammar in the software interface documentation. A natural interaction can be established.

[0017] Preferably, before the language module outputs the machine grammar, the module requests confirmation whether the machine grammar is the correct translation of the input query.

[0018] A method, in particular a computer-implemented method, for generating prompts and / or training datasets for large language models, in particular for using the methods described herein; comprising: - Establishing a connection between an input query, in particular an input query in natural language, and a machine grammar, preferably a machine grammar derived from a document.

[0019] A prompt and / or training dataset for use in the method for training a large language model described herein; the method comprising: - Preferably, a plurality of example input queries expressed in natural language, and translations from the input queries to machine grammar, and / or - Preferably, a task description that specifies, in natural language, the expected behavior of the language module with respect to its output and a specific format of the machine grammar that may be included therein.

[0020] Furthermore, it is proposed to use the methods disclosed herein for communication with household appliances.

[0021] A data processing system comprising means for performing the methods described herein.

[0022] Preferably, the language module resides on a high-performance computer, preferably in the cloud, which is also often referred to as a server. Preferably, the heavy workload of the language module is outsourced to the server, while preferably, the computing unit, in particular the computing unit embedded in the household appliance, can be relatively small, especially compared to the server. Alternatively, a distilled version of the large language model is hosted on the embedded computing unit, where in particular, the computing performance is lower than that of the server. In particular, the server can be hosted in the company's infrastructure. Preferably, the distilled version is hosted on the household appliance.

[0023] A computer program comprising instructions that, when executed by a computer, cause the computer to perform the methods described herein.

[0024] A computer-readable medium comprising instructions that, when executed by a computer, cause the computer to perform the methods described herein.

[0025] The method can be implemented as a computer program product with a program code device and can be stored on a computer-readable medium. The method can be adapted to be executed, in whole or in part, by a processing device. The processing device can be of an electronic nature and include a microcomputer or a microcontroller, an ASIC, or a similar device. The features or advantages of the method can apply to the corresponding device or system, and vice versa. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The present invention will now be discussed in more detail with reference to the drawings, in which: Figure 1 A schematic illustration of the communication between a user and a household appliance is shown; Figure 2 A schematic and exemplary illustration of a language module for communication is shown; Figure 3 A flowchart of a method for controlling a household appliance is shown; Figure 4 A flowchart of a method is shown; Figure 5 A flowchart of a method for prompting and / or training a large language model is shown; Figure 6 A prompt and / or training dataset, a processing system, a workstation, a computer program, and a computer-readable medium are shown. DETAILED DESCRIPTION

[0027] Figure 1 User 100 is shown, and user 100 communicates with household appliance 105. In a broader sense, household appliance 105 can be any Internet of Things (IoT) device. Household appliance 105 can be, by way of example only, a washing machine, a dryer, a dishwasher, a stove, a refrigerator, a freezer, a vacuum cleaner, or any household appliance. User 100 sends input query 110 to household appliance 105, particularly a task-based query 115. Such a task-based query 115 can be, for example, to configure a washing machine, particularly to operate the washing machine and / or make a request for the washing machine, by a program name (such as, for example, "cotton fabric", "mixed fabric", "silk", etc.) and / or specifications (such as "temperature "20", "30", "90", etc., by way of example only). Another example of a task-based query is "how to remove a grass stain on a silk shirt". The message, particularly input query 110, can also be and / or include an auxiliary query 116, such as a question about household appliance 105, such as, for example, "how to clean the washing machine".

[0028] The household appliance 105 sends an output 120 to the user 100. The output 120 is preferably an information message in response to an input query 110. If the input query 110 is a task-based query 115, the output 120 can be, for example, "Confirmed, the program is for cotton fabrics and the temperature is 90 °C". If the input query 110 is an auxiliary query 116, the output 120 can be, for example, "To treat grass stains on a silk shirt, try applying some mild detergent directly to the stain" or "To clean the washing machine, add 100 ml of washing machine cleaner to the detergent tray". The input query 110 can also include a task-based query 115 and an auxiliary query 116. For example, if the input query 110 is "I put a silk shirt with an oil stain into the washing machine", the task-based query 115 can be to set the silk washing program, and the auxiliary query 116 can be to answer "I recommend that you start the silk washing program, which will help clean your silk shirt, including the oil stain. The program will take a little time so that all the oil stains can be cleaned. If you wish, I can start the program now". These are just some examples, and many other examples are conceivable.

[0029] The input query 110, in particular the task-based query 115, in particular the auxiliary query 116, and the output 120 are preferably voice messages and voice outputs. In a further embodiment, the input query 110 and the output 120 can also be any kind of message, such as a text message, a visual message, etc.

[0030] Figure 2 A schematic and exemplary illustration of a language module 200 for communication, in particular for communication between the user 100 and the household appliance 105, is shown. The language module 200 includes a dialogue manager 205. The language module 200 includes a word detector 210, in particular a hotword and / or wake word detector. Once the user 100 sends an input query 110, in particular a task-based query 115 and / or an auxiliary query 116, the word detector 210 detects the message. After detecting the input query 110, the word detector 210 sends 215 the detected input query 110 to the dialogue manager 205. The dialogue manager 205 sends 220 back to the word detector 210 that the input query 110 has been received and particularly instructs the word detector 210 to wait. The dialogue manager 205 sends 225 the input query 110 to an automatic speech recognition device 230, particularly for the case when the input query 110 is a voice message. The automatic speech recognition device 230 translates the voice input query 110 into a text input query 110. The automatic speech recognition device 230 sends 235 the input query 110, particularly as a text input query 110, back to the dialogue manager 205.

[0031] The dialogue manager 205 sends 240 the input query 110, especially as a text input query 110, to the large language model device 245. The large language model device 245 is, for example, a device that includes a large language model, especially a machine learning model. The large language model determines whether the input query 110 includes a task-based query 115 and / or an auxiliary query 116.

[0032] If a task-based query 115 is detected, the language module 200, especially the large language model device 245, translates the task-based query 115 into a machine grammar. The machine grammar is a grammar that can be understood by the household appliance 105. The large language model device 245 sends 250 the machine grammar to the dialogue manager 205, and the dialogue manager 205 sends 255 the machine grammar to the household appliance 105. The household appliance 105 executes the machine grammar and confirms 260 to the dialogue manager that the machine grammar has been executed. The dialogue manager 205 sends 265 a confirmation message to the text-to-speech device 270. The text-to-speech device 270 converts the confirmation message into a voice message and sends 275 the voice message to the dialogue manager 205. The dialogue manager 205 notifies the user 100 via the voice message that the machine grammar has been executed.

[0033] If an auxiliary query 116 is detected, the language module 200, especially the large language model device 245, finds an answer to the auxiliary query 116. The large language model device 245 sends 250 the answer to the dialogue manager 205. The dialogue manager 205 sends 265 the answer to the text-to-speech device 270. The text-to-speech device 270 converts the answer into a voice answer and sends 275 the voice answer to the dialogue manager 205. The dialogue manager 205 notifies the user 100 of the voice answer.

[0034] The input query 110 can include a task-based query 115 and an auxiliary query 116. This can even be used for short input queries 110. For example, "I have put the dirty cotton fabric into the machine" can include a task-based query 115 to start a program for cleaning the cotton fabric, and also answer the user about which program is recommended and / or give any recommendations to the user about "dirty cotton fabric" and / or answer.

[0035] Figure 3 An exemplary method 300 for controlling a household appliance 105 using a language module 200, which includes a large language model, especially a large language model device 245, is shown, and the method includes: - The language module 200 receives 300 the input query 110; - The language module 200 determines 305 whether the input query 110 includes a task-based query 115; - The language module 200 translates 310 the task-based query 115 into a machine grammar that can be understood by the household appliance 105; - The language module 200 outputs 315 the machine grammar; - The household appliance 105 receives 320 the machine grammar; and - The household appliance 105 executes 325 the machine grammar.

[0036] Figure 4 Exemplarily shown is a method 400 for prompting and / or training a large language model to control a household appliance 105 using the language module 200, which includes a large language model, particularly the large language model device 245, and is particularly for using the methods described herein; the method 400 includes: - Receiving 405 an input prompt and / or training dataset including a connection between an input query and a machine grammar, particularly where the input prompt and / or training dataset is derived from a manual; - Particularly, where before the language module 200 outputs the machine grammar, the module requests 410 confirmation as to whether the machine grammar is a correct translation of the input query.

[0037] Figure 5 Exemplarily shown is a method 500 for generating a prompt and / or training dataset for a large language model, particularly for using the methods described herein; including: - Establishing 505 a connection between the input query 110 and the machine grammar.

[0038] Figure 6 Exemplarily shown is a prompt and / or training dataset 600 for use in the methods of training a large language model described herein; the method includes: - An input query 110 expressed in natural language; - A translation from the input query 110 to the machine grammar.

[0039] Preferably, the methods described herein are for communication with the household appliance 105.

[0040] Figure 6 Shown is a data processing system 605 that includes means for performing the methods described herein. The data processing system 605 preferably includes a language module that preferably resides on a workstation 610. The workstation 610 is particularly a powerful workstation. The workstation 610 can be, for example, in the cloud, where the household appliance 105 has a connection to the cloud, preferably an Internet connection.

[0041] Figure 6Exemplarily shown is a computer program 615 including instructions which, when executed by a computer, cause the computer to perform the methods described herein. Figure 6 Exemplarily shown is a computer-readable medium 620 including instructions which, when executed by a computer, cause the computer to perform the methods described herein.

[0042] The methods described herein and the systems described herein are preferably used to customize the behavior of large language models, particularly focusing on their flexible ability to both provide natural language answers to user questions and generate messages in a specific syntax (particularly machine syntax) so as to preferably control household appliances 105 and / or particularly other Internet of Things (IoT) devices via an application programming interface (API). Large language models (such as, for example, GPT 3.5) are designed to understand and generate human-like text based on a vast amount of pre-trained knowledge. The methods described herein utilize the potential of large language models to create customized responses covering two different applications, particularly first (preferably via voice) controlling and / or manipulating household appliances 105 and particularly second obtaining domain-specific voice assistance in the field of household appliances.

[0043] Large language models preferably operate by adopting a neural network architecture based on a transformer-based model which preferably utilizes a vast amount of pre-existing text data for context understanding and text generation. In particular, through training on a large number of data sources, large language models have acquired proficiency in understanding and generating human-like text.

[0044] In particular, to provide natural language answers to user questions, the system preferably utilizes the ability of the large language model to comprehend input prompts, context, and / or user queries, particularly input query 110. Preferably, this is achieved by fine-tuning and customizing model parameters to fit a specific domain or task. In particular, user 100 can input a question, query, or request, and the system generates a coherent, contextually relevant response.

[0045] In particular, in the context of preferably controlling household appliances 105 and / or IoT devices via an API, the system utilizes the text generation ability of the language model to produce messages in a specific syntax understood by household appliances 105, and / or IoT devices, and control systems. Preferably, the syntax includes commands, as well as any parameters and data required to interact with the device (particularly household appliance 105) for remote control. Preferably, by providing instructions and parameters to the model, the large language model produces a customized message that allows seamless communication with household appliance 105.

[0046] In particular, messages in a specific syntax for controlling household appliances and / or any IoT devices (preferably via an API) are first generated, and in particular, secondly, natural language answers to user assistance queries in the same household appliance domain can also be provided.

[0047] Preferably, efficient and intuitive control of the household appliance 105 (especially via voice) is achieved, as well as generating an expert-level assisted response to user questions posed by the user via voice.

[0048] In one embodiment of the present invention, technically, large language models (such as general large language models, e.g., GPT-3.5) are customized via a technique called "prompting". In particular, the concept of large language model prompting revolves around the ability to interact with these complex large language models by providing them with specific instructions (called (input) prompts). Preferably, this process enables the customization of the behavior of large language models for various applications. For example, when a prompt is input, the large language model processes the text and generates a response based on its vast pre-trained knowledge and understanding of the provided context. Preferably, by carefully crafting the prompt, the desired output for two applications (especially task-based queries and assistance queries) can be elicited.

[0049] In particular, the large language model is input as a new software module into the language module, which, in particular, typically resides on a powerful computer (e.g., a workstation) in the cloud. Preferably, this computer also has a software interface compatible with protocols known from the prior art (e.g., the Hermes protocol), i.e., it can "subscribe to" (read) specific protocol messages originating from other software modules and "publish" (write) other specific protocol messages that will be read by other software modules further down in a typical message chain.

[0050] Preferably, the large language model includes an input query and an output query.

[0051] The input query can occur, for example, as specified in a protocol (e.g., the Hermes protocol), with additional data, especially these additional data are wrapped in a so-called "message payload", which, for example, contains a user command, especially a user command that the ASR component has previously transcribed from audio into text form. Through this type of message, the large language model module will particularly receive the above two types of queries, especially task-based queries and assistance queries.

[0052] The large language model module preferably provides one or more of the following message types, particularly depending on the internal decision logic of the large language model, which is preferably specified by instructing the model via a prompt: Specifically, first, for task-based queries, especially in cases where the user expects the state of a household appliance to be changed / manipulated as a result of their query (voice command). Specifically, second, when the payload of these messages carries natural language text, this text will subsequently be converted to audio by the text-to-speech module and specifically played to the user via a speaker. These messages can be subdivided as follows: First, specifically, for task-based queries, in cases where the large language model has generated an output, the large language model supplements the output with a meaningful and typically short natural language message that will confirm to the user the change in the household appliance state, e.g., "I have selected the mixed fabric program for you and added the stain removal option." Second, specifically, for auxiliary queries, the large language model is instructed (specifically via a prompt) to grammatically wrap its "normal" text response, e.g., wrap it into the payload of a "hermes / tts / say" message.

[0053] Advantageously and preferably, this technical solution allows for the removal of some or all of the "voice responses of statically stored dialogue systems" known from the prior art, particularly those previously encoded in "client code", and particularly in cases where, for example, the client code processes "hermes / intent / <intentname>Those responses from which the "message" was retrieved when.

[0054] Regarding any previous natural language understanding solutions, particularly rule - based statistical natural language understanding solutions for task - based speech use cases, the benefits of the extremely superior natural language understanding (NLU) performance brought by large language model technology can be shown. In particular, in the context of natural language understanding, "performance" (also known as "robustness") describes the ability of a system to understand speech commands regardless of how the user's question is phrased, particularly regarding first lexical (e.g., as in words / synonyms) selection and variation; particularly regarding syntactic (e.g., as in sentence structure) selection and variation.

[0055] The benefits of the vast amount of domain - independent / general "world knowledge" captured in a large language model can be shown. The large language model may be enhanced, particularly by a "two - step method", to utilize domain - specific knowledge. Preferably, the voice control system can be technically docked with a host IoT device and / or household appliances to read or modify the state of the device and / or household appliances.

[0056] As an example, an exemplary answer to an exemplary question will be described, where the answer includes the user's natural language answer and machine grammar: The user asks: "How do I wash my cotton scarf?" The language module 200 outputs an answer to an auxiliary query: "I have selected the most suitable program for you: the cotton fabric program. This program is specifically designed for cleaning cotton or linen clothes with normal to heavy soiling. The maximum load of this program is 5 kg. Please add your scarf to the drum and close the door." The language module 200 outputs machine grammar: "intent:configure_washer_program; programKey:LaundryCare.Washer.Program.Cotton". Preferably, the "intent" value describes the potential intent of the user's command; preferably, the "programKey" is a slot whose value identifies the machine - readable washer program selected by the user's intent.

[0057] Reference list (as part of the specification) 100 User 105 Household appliances 110 Input query 115 Task - based query 116 Auxiliary query 120 Output 200 Language module 205 Dialogue manager 210 Word detector 215 Transmission 220 Transmission 225 Transmission 230 Automatic Speech Recognition Device 235 Transmission 240 Transmission 245 Large Language Model Device 250 Transmission 255 Transmission 260 Confirmation 265 Transmission 270 Text-to-Speech Device 275 Transmission 300 Method 305 Judgment 310 Translation 315 Output 320 Reception 325 Execution 400 Method 405 Reception 410 Request 500 Method 505 Establishment 600 Training Dataset 605 Processing System 610 Workstation 615 Computer Program 620 Computer-readable Medium< / intentname>

Claims

1. A method (300) for controlling a household appliance (105) using a language module (200) including a large language model, the method (300) comprising: - The language module (200) receives (300) an input query (110); - The language module (200) determines (305) whether the input query (110) includes a task-based query (115); - The language module (200) translates (310) the task-based query (115) into a machine grammar, where the machine grammar is understandable by the household appliance (105); - The language module (200) outputs (315) the machine grammar; - The household appliance (105) receives (320) the machine grammar; and - The household appliance (105) executes (325) the machine grammar.

2. The method (300) according to claim 1, wherein what the language module (200) outputs is: The machine grammar has been sent to the household appliance (105).

3. The method (300) according to claim 2, wherein the output (120) is converted into speech, in particular a speech output.

4. The method (300) according to one of the preceding claims, wherein the language module (200) determines whether the input query (110) includes an auxiliary query (116); and wherein the language module (200) outputs an answer to the auxiliary query (116).

5. The method (300) according to one of the preceding claims, wherein the household appliance (105) is voice-controlled using voice commands; wherein the voice commands are converted into an input query (110); wherein the input query (110) is in text form.

6. A method (400) for prompting and / or training a large language model to control a household appliance (105) using a language module (200) including a large language model, in particular for using the method (300) according to one of the preceding claims; the method (400) comprising: - Receiving (405) an input prompt and / or training data set (600) including a connection between an input query (110) and a machine grammar.

7. The method (400) according to claim 6, wherein the input prompt and / or training data set (600) is derived from a manual.

8. The method (400) according to claim 6 or 7, wherein before the language module (200) outputs the machine grammar, the language module (200) requests confirmation whether the machine grammar is a correct translation of the input query (110).

9. A method (500) for generating a prompt and / or training data set (600) for a large language model, in particular for using the method (300) according to one of claims 1 to 5; the method (500) comprising: - Establishing (505) a connection between an input query (100) and a machine grammar.

10. A prompt and / or training data set (600) for use in the method (500) of training a large language model according to claim 9, comprising: - An input query (110) expressed in natural language; - Translation from an input query (110) to a machine grammar.

11. Use of the method (300) according to any one of claims 1 to 5 for: - Communication with a household appliance.

12. A data processing system (605) comprising means for performing the method (300, 400, 500) according to any one of claims 1 to 9.

13. The data processing system (605) according to claim 12, wherein the language module (200) resides on a workstation (610).

14. A computer program (615) comprising instructions which, when executed by a computer, cause the computer to perform the method (300, 400, 500) according to any one of claims 1 to 9.

15. A computer-readable medium (620) comprising instructions which, when executed by a computer, cause the computer to perform the method (300, 400, 500) according to any one of claims 1 to 9.

Images