LLM-based question and answer method

Abstract

The present invention relates to a computer-implemented question and answer method which is operated via an API by means of an LLM (large language model)-based chatbot system (4) and specified by the LLM-based chatbot system (4) within the framework of the question and answer method, the following participants interact by means of messages: a first participant (T1) in the form of the LLM (large language model)-based chatbot system (4), which is designed to complete messages in response to an instruction, wherein a SYSTEM attribute is assigned to the first participant (T1); a second participant (T2) to which a message is to be communicated within the framework of the question and answer method in response to a question in the question and answer method, wherein a USER attribute is assigned to the second participant (T2); a third participant (T3), which performs a chatbot assistant function and sends a message to the first participant (T1) in order to trigger a message for the first participant (T1), wherein a CHATBOT attribute is assigned to the third participant (T3), wherein in a departure from the role assignment specified by the LLM-based chatbot, the roles of the attributes are switched between the USER attribute and the CHATBOT attribute such that the answer, generated by the chatbot system (4), to a question from the third participant (T3) with the switched CHATBOT attribute generates a chatbot answer oriented to the CHATBOT attribute, not the USER attribute.

Description

[0001] LLM-based question and answer procedure

[0002] The present invention relates to an LLM-based question and answer method according to the preamble of claim 1. The LLM-based question and answer method is particularly intended for chatbot-based communication between a user (customer) and an online shop.

[0003] Technological background

[0004] The present invention relates, in its basic principle, to a computer-implemented question and answer method for providing an answer in the form of a message, in particular a voice or text message, to a user's question using a so-called LLM ("Large Language Model") model, in which a chat with a chatbot is operated via a Chat Completion API or a chat version of an LLM. The following participants interact here: A first participant in the form of an LLM-based chatbot system that is configured to complete messages upon instruction, with a system attribute assigned to the first participant. The first participant can access one or more LLM-based databases of the chatbot system.The second participant is a user who, as part of the question and answer process, is to be sent a message in response to a question from the user. A USER attribute is assigned to the second participant. Furthermore, a third participant is provided, which performs a chatbot assistant function and sends a message to the first participant in order to trigger a message to the first participant. A CHATBOT attribute is assigned to the third participant. Such question and answer processes are based on LLM databases and the use of neural networks. Corresponding processes open up a wide range of applications, such as telephone assistance processes, chatbots, customer service, e-banking, product consulting, and the like.

[0005] Printed state of the art

[0006] US Pat. No. 10,664,527 B1 discloses an AI-based question and answer procedure designed to improve the quality of communication between a user and the system. The procedure uses a pre-trained model to identify answers and subsequently generate answer vectors. The answer vectors are associated with context vectors, with answer vectors being generated based on a weighted context vector.

[0007] WO 2019 / 118254 A1 describes a chatbot that integrates a derived user intent as a function. The method comprises receiving a message from a user and generating an intent vector by processing the user input using an AI model that has been trained with data representing the user intent. An intent vector comprises a probability for each intent in a known number of possible intents.

[0008] Previous question and answer systems suffer from the problem that after a certain chat history, the user only receives poor-quality message suggestions. For example, if the user is interested in a particular product category in an online shop, they no longer receive truly useful information as the chat history progresses.

[0009] Object of the present invention

[0010] The object of the present invention is to improve the chat flow with regard to the information exchange desired by the user. Solution to the problem

[0011] The above object is achieved by the computer-implemented question and answer method according to claim 1 and the independent claim 8. Advantageous embodiments of the method according to the invention are claimed in the dependent claims.

[0012] The special feature of the method according to the invention is that, in contrast to the role assignment predetermined by the LLM (Large Language Model)-based chatbot system, the roles of the attributes are swapped between the USER attribute and the CHATBOT attribute, so that the response generated by the chatbot system to a question from a third participant under the swapped CHATBOT attribute generates a CHATBOT response oriented towards the CHATBOT attribute and not the USER attribute. By means of such a newly constructed chat history with swapped roles, qualitatively better text suggestions (e.g. several concrete suggestions to choose from) can be generated, even when using relatively simply designed chatbot systems such as the GPT-3.5® chatbot system from OpenAI®. This is because chatbot systems are generally programmed in such a way that the messages directed by the system to a user are designed to inform the user (e.g.customers of an online shop).

[0013] Ideally, the role reversal is already established by the initial prompt instruction to the first participant at the start of the question and answer process. A "prompt" is an instruction to the chatbot system or its LLM regarding what result the chatbot system or the associated LLM should produce.

[0014] According to a further embodiment of the present invention, the role reversal is hidden from the second participant (i.e., the user). The significance of the role reversal plausibly arises from the fact that, with regard to the LLM-based chatbot system used, after an identical chat history with identical instructions (prompt) to the first participant, a chatbot response oriented towards the CHATBOT attribute generates a more specific message than a CHATBOT response oriented towards the USER attribute. For example, instead of advice, three selection suggestions (e.g., product suggestions) are communicated. This difference is due to the fact that the algorithm of the chatbot system or its LLMs, with regard to its vectorization, is geared towards presenting the user with more general information in a conventional role distribution.

[0015] According to the method according to the invention, it can be provided that an instruction (prompt) is sent to the first participant in the chat conversation, which is intended to generate the response based on the CHATBOT attribute rather than the USER attribute in the form of several suggested answers (e.g., product features). The suggested answers enable the third participant or user to generate a selection.

[0016] Within the scope of the method according to the invention, the history of the respective chat process can be transmitted again with each voice or text message.

[0017] The response, which is based on the CHATBOT attribute rather than the USER attribute, can be an audio file, text file, image file, or a combination thereof.

[0018] The above object is furthermore also achieved by a generic method which is also claimed in a co-dependent manner according to claim 8. In this case, the chat process is divided into several phases, wherein in a first phase the second participant or user posts a message in the chat, in a second phase the second participant receives a message which includes several options for the second participant to choose from, and in a third phase the second participant selects one of the several options, wherein the chat process with regard to the messages to the second participant is supplemented by further message elements (e.g. incentives to complete a purchase) after the selection has been made. This creates the advantage that the second participant or user is not only advised by the chatbot system, but is also guided to a destination using the chatbot system with regard to a message submitted by the second participant or user.User expected selection or purchase decision can be made.

[0019] According to the invention, during the chat, it can be determined on the basis of the chatbot system or its LLMs which of the phases is currently present in the chat.

[0020] Furthermore, content elements associated with the second participant or user (i.e., those communicated by them in the chat) can be defined, which at least include the following: a request for more detailed information, e.g., about a product and / or a technical question, and / or a request for a product comparison, and / or a restart of a query, and / or an above-average chat duration. During the chat, the chatbot or its LLM-based database determines which content element is currently present.

[0021] According to the invention, during the chat, the chatbot system or its LLMs can be used to determine which of the elements is currently present in the chat.

[0022] Conveniently, an analytical implementation of the assignment of the first, second, or third phase to the content element belonging to the second participant or user can be carried out on the basis of the chatbot or its LLM-based database, and as a result, a message, e.g. in the form of an incentive, can be integrated into the chat for the second participant. Compared to previous methods, this can result in improved motivation, e.g. for a purchase decision, by the second participant or user. According to one embodiment, the aforementioned message, ie the incentive message, can be inserted into the current chat history via a prompt from the content of a predefined table with multiple message options.

[0023] The table can be continuously updated using a machine learning process.

[0024] The update is preferably based on previous chats and / or resulting purchase decisions.

[0025] Description of the invention using an embodiment

[0026] A suitable embodiment of the method according to the invention is described in detail below with reference to the drawings. Identical features are provided with a reference numeral only once for clarity. They show:

[0027] Fig. 1 is a schematic representation of an example of a chatbot system for applying the method according to the invention;

[0028] Fig. 2 shows a schematic representation of the participants of the chatbot-based question and answer procedure according to the invention and the role reversal;

[0029] Fig. 3 is a flowchart showing an example of a conventional chatbot-based chat flow;

[0030] Fig. 4 shows a flowchart showing an example of a chat flow based on role reversal according to the invention; Fig. 5 shows a flowchart showing an example of a sequence of the question and answer method according to the invention for improving the targeting of a chatbot-based chat;

[0031] Fig. 6 a schematic representation of a definition of phases of the chat process, a definition of elements of the chat process and a resulting definition of options for incentive information to the user as well as

[0032] Fig. 7 shows an example of a chat flow in which incentive information is generated based on the defined phases and elements.

[0033] The method according to the invention is a computer-implemented question and answer method for providing an answer in the form of a message, in particular a voice or text message, to a user's question using an LLM (Large Language Model). Such models are based on the use of artificial intelligence and allow the operation of a chatbot system 4 shown in Fig. 1, which comprises an LLM-based neural network e, an input option ?, and an output option 8 to the network 6.

[0034] The chatbot system 4 is designed to complete messages to a user upon instruction. The chatbot system 4 accesses data from at least one database 5 for this purpose. A plurality of databases can also interact with the chatbot system 4 for access. The chatbot system 4 is accessible for connection via a so-called Chat Completion API. A "Chat Completion API" is an API (Application Programming Interface) that enables developers to interact with the chatbot system 4. The LLM is provided as a cloud service. Alternatively, the LLM can also be provided as a chat version of an LLM, i.e., as a local installation ("Chat Model"). The chatbot system 4 can, for example, be used by an online shop 13 for an automated question and answer process with a user T2 (e.g., customer). The user T2 can use a terminal device 1 via a publicly accessible network 2 (e.g.,The chatbot system 4 can communicate bidirectionally with the online shop 13 (e.g., via the Internet). Furthermore, network devices 3 can be provided. The end device on the user T2 side can be a PC, a smartphone, or a tablet. The chatbot system 4 can be, for example, the ChatGPT® chatbot system, e.g., version 3.5 from OpenAI®.

[0035] According to Fig. 2, the chatbot system 4 is configured for multiple participants, namely a first participant T1 in the form of the chatbot system 4. The first participant T1 is assigned the SYSTEM attribute. In addition, there is a second participant T2, to whom a message is to be communicated in response to a question in the question and answer process. The second participant T2 is assigned a USER attribute. This could be, for example, a potential customer or prospective buyer of the online shop 13 shown in Fig. 1. Furthermore, a third participant T3 is defined, who performs a chatbot assistant function in the chatbot system 4 and sends a prompt to the first participant T1 to trigger a message from the first participant T1. The third participant T3 is assigned a CHATBOT attribute. This is the typical structure of role distribution in a chatbot system 4.

[0036] According to the present invention, in the chatbot system 4, as indicated by the two arrows, the roles of the attributes are swapped between the USER attribute and the CHATBOT attribute, in contrast to the role assignment predetermined by the LLM-based chatbot, so that the response generated by the chatbot system 4 to a question from the third participant T3 under the swapped CHATBOT attribute generates a chatbot response oriented towards the CHATBOT attribute and not the USER attribute. This ensures that the system outputs high-quality text suggestions to the user, since the chatbot system 4 no longer has to generate general messages due to the role swap regarding its vectorization, but can instead generate specific messages. Due to the role swap, the LLM's programming, according to which the LLM should only advise the customer and not generate messages that the user should actually provide, is thus bypassed.This avoids bad message suggestions from the LLM and enables an improved chat flow.

[0037] The role reversal is preferably determined by an initial instruction (prompt) to start the question and answer process to the system, i.e., to the first participant T1. The role reversal remains hidden from the user, i.e., the second participant T2. With each voice or text message, the entire chat history is retransmitted. A message or the response, which is based on the CHATBOT attribute rather than the USER attribute, can be an audio file, text file, or image file, or a combination thereof.

[0038] Figures 3 and 4 show, by way of example, the fundamental difference in the chat flow between the previous method without role reversal, as shown in Figure 3, and the method according to the invention with role reversal as shown in Figure 4. The comparison clearly shows that, due to the role reversal according to the invention, the chat flow according to Figure 4 can be significantly more goal-oriented than the chat flow according to Figure 3. For example, due to the programming of the LLM, the system, i.e., participant T1, can generate three selection suggestions, e.g., USB sticks, in response to the message from user T2, whereas in the chat flow according to Figure 3, only another general response can be generated.

[0039] The newly constructed chat flow with swapped roles, according to the invention, can thus generate high-quality text suggestions, even when using comparatively simple LLMs or chatbot systems, such as the product ChatGPT-3.5.® from OpenAI®. Fig. 5 shows another option for improving the chat flow or motivating a purchase decision in a chatbot-assisted online shop. The chat flow is divided into several phases P1-P3 (step 100 in Fig. 5). These phases can be, for example, the phases shown above in Fig. 6, namely:

[0040] P1 : the customer has communicated his requirements,

[0041] P2: suitable products were suggested, P3: the customer selected a product.

[0042] According to step 101 of Fig. 5, the chat history is continuously examined for the presence of predefined elements of user behavior. Corresponding elements of user behavior can be defined in advance and are shown as examples in Fig. 6. Corresponding elements can be:

[0043] E1 : User wants more detailed information about a product,

[0044] E2: User has a technical question,

[0045] E3: User wants to compare products,

[0046] E4: User changes his mind and restarts,

[0047] E5: Chat takes longer than average.

[0048] According to step 102 of Fig. 5, the chatbot system 4 or its LLM determines during the chat which phase and which element are currently present.

[0049] According to step 103, the chatbot system 4 or its LLM searches a fixed table (incentive information table) to determine which information should be integrated into the chat as an incentive. The selection is made through an analytical implementation of the assignment of the respective phases P1-P3 to the respective elements E1-E5. The information selected by the LLM is integrated into the chat history via a prompt according to step 104 in Fig. 5. This is preferably incentive information, which, according to Fig. 6, can be configured as follows: I) the chatbot points out product ratings, e.g., a "very good" rating,

[0050] II) the chatbot indicates the current stock level of the product,

[0051] III) the chatbot offers a voucher,

[0052] IV) the chatbot avoids mentioning new products.

[0053] According to step 105 of Fig. 5, the table in which the relevant incentive information is stored is continuously updated by the LLM through machine learning, based on the previous chats and the resulting user decisions.

[0054] Fig. 7 shows an exemplary chat flow using the previously described analytical implementation of assigning a predefined phase P1-P3 and a predefined element E1-E5 to create a purchase incentive. The chat flow spans the three phases P1 to P3, as shown in Fig. 6 above, with the LLM analyzing which elements E1-E5 are present depending on user behavior. For example, the user has selected a product based on several product suggestions from the assistant (chatbot assistant) and would like more detailed information about this product (phase P3 in Fig. 7). This is followed by a corresponding prompt, after which the system sends an incentive message, e.g., in the form of a note that the stock of the suggested product is low.

[0055] The method according to the invention enables a significant increase in the efficiency of an automated question and answer process using a chatbot system. It therefore represents an important contribution to the relevant field of technology. REFERENCE SYMBOL LIST

[0056] 1 device (user)

[0057] 2 Network

[0058] 3 Network setup

[0059] 4 Chatbot system

[0060] 5 Database

[0061] 6 LLM based neural network

[0062] 7 Input

[0063] 8 Output

[0064] 9 Chat Completion API

[0065] 10 systems

[0066] 11 users

[0067] 12 Assistant

[0068] 13 Online Shop

[0069] T1 system

[0070] T2 User

[0071] T3 Assistant

Claims

PATE NTAN SPRÜ CH E 1. A computer-implemented question and answer process for providing an answer in the form of a message, in particular a voice or text message, to a user's question using an LLM (Large Language Model) model, wherein the question and answer process is operated via an API using a chatbot system (4) based on LLM (Large Language Model), and within the framework of the question and answer process, the following participants interact via messages as specified by the LLM-based chatbot system (4): a first participant (T1) in the form of the chatbot system (4) based on LLM (Large Language Model), which is configured to complete messages upon instruction, wherein the first participant (T1) is assigned a SYSTEM attribute, a second participant (T2), to whom a message is to be communicated within the framework of the question and answer process in response to a question in the question and answer process,wherein the second participant (T2) is assigned a USER attribute, a third participant (T3) who has a chatbot assistant function and sends a message to the first participant (T1) to trigger a message to the first participant (T1), wherein the third participant (T3) is assigned a CHATBOT attribute, characterized in that, in deviation from the role assignment predetermined by the LLM-based chatbot, the roles of the attributes between the USER attribute and the CHATBOT attribute are swapped, so that the response generated by the chatbot system (4) to a question from the third participant (T3) under the swapped CHATBOT attribute generates a chatbot response oriented towards the CHATBOT attribute and not the USER attribute.

2. Question and answer procedure according to claim 1, characterized in that the role reversal is determined by the initial instruction (prompt) to start the question and answer procedure to the first participant (T1).

3. Question and answer procedure according to claim 1 or 2, characterized in that the role reversal is hidden from the second participant (T2).

4. Question and answer method according to at least one of the preceding claims, characterized in that the chatbot system (4) based on the LLM used generates a more specific message after an identical chat history with an identical instruction (prompt) to the first participant (T1) in comparison to a chatbot response oriented to the USER attribute.

5. Question and answer method according to at least one of the preceding claims, characterized in that in the chat history an instruction (prompt) is given to the first participant (T1) which aims to generate the answer oriented towards the CHATBOT attribute and not the USER attribute in the form of several suggested answers.

6. Question and answer procedure according to at least one of the preceding claims, characterized in that the history of the chat process is transmitted again with each voice or text response message.

7. Question and answer procedure according to at least one of the preceding claims, characterized in that the response oriented towards the CHATBOT attribute and not the USER attribute is an audio file, text file and / or image file.

8. A computer-implemented question and answer process for providing an answer in the form of a message, in particular a voice or text message, to a user's question using a GPT (Generative Pre-trained Transformer) model, wherein the question and answer process is operated by means of a chatbot system (4) based on LLM (Large Language Model), and within the framework of the question and answer process, the following participants interact via messages as specified by the LLM-based chatbot system (4): a first participant (T1) in the form of the LLM (Large Language Model)-based chatbot system (4), which is configured to complete messages upon instruction, wherein the first participant (T1) is assigned a SYSTEM attribute, a second participant (T2), to whom a message is to be communicated within the framework of the question and answer process in response to a question in the question and answer process,wherein the second participant (T2) is assigned a USER attribute, a third participant (T3) who has a chatbot assistant function and sends a message to the first participant (T1) in order to trigger a message to the first participant (T1), wherein the third participant (T3) is assigned a CHATBOT attribute, in particular according to at least one of the preceding claims, characterized in that the chat history is divided into several phases (P1, P2, P3), it is determined in which phase the chat history is, and the chat history is supplemented with further message elements (NE) with regard to the messages from the first participant (T1) to the second participant (T2) after the selection has been made.

9. Question and answer method according to claim 8, characterized in that a first, second and third phase (P1, P2, P3) are defined, wherein in the first phase (P1) the user communicates his requirements, in the second phase (P2) messages, e.g. in the form of product suggestions, are communicated to the user, in the third phase (P3) the user selects or has selected a message, e.g. a product, and during the chat it is determined on the basis of an LLM-based database which of the phases (P1, P2 or P3) are currently present.

10. Question and answer method according to claim 8 or 9, characterized in that content elements (E1 - E1+n) associated with the second participant (T2) are defined, which preferably comprise the following: a request for more detailed information on a product (first content element E1), and / or a technical question (second content element E2), and / or the request for a product comparison (third content element E3), and / or a restart of the query (fourth content element E4), and / or an above-average duration of the chat (fifth content element E5).

11. Question and answer method according to claim 9, characterized in that an analytical implementation of the assignment of the phases (P1 - P3) to the content elements (E1 - E1+n) belonging to the second participant (T2) takes place and as a result thereof a message to the second participant (T2) is integrated into the chat.

12. Question and answer procedure according to claims 8 to 11, characterized in that from the content of a fixed predetermined Table with several possible messages, one of which is inserted into the chat history via a prompt.

13. Question and answer method according to claim 12, characterized in that the table is continuously updated by a machine learning method.

14. Question and answer procedure according to claim 13, characterized in that the update is based on previous chats and / or resulting purchase decisions.

Images