System for answering questions

Abstract

A system for answering questions is proposed, wherein the system comprises a receiving unit for receiving a question from a user and a processing unit for generating an answer to the question of the user using artificial intelligence and for outputting the answer, wherein the answer has a predefined confidence score, the receiving unit is configured to receive feedback from the user on the outputted answer, and the processing unit is configured to adjust the predefined confidence score based on the feedback.

Description

[0001] 24-2858 PIF

[0002] System for answering questions

[0003] Description

[0004] The present invention relates to a system for answering questions according to the preamble of claim 1. Furthermore, the present invention relates to a method for answering questions according to the preamble of claim 9.

[0005] Systems based on or utilizing artificial intelligence are becoming increasingly widespread. One example is generative artificial intelligence (hereinafter also referred to as generative AI). This is a form of artificial intelligence that is trained on large datasets (from existing (training) data) to create new content such as texts, images, or videos. In particular, such systems use neural networks. However, existing generative AI systems do not sufficiently consider how a user interacts with the data provided by the AI ​​system, especially whether a user accepts this data or asks further questions of the AI ​​system based on the provided data.

[0006] Current generative AI systems, such as chatbots or image generators, are based on algorithms trained on large datasets to respond to specific user input. However, these systems are primarily designed to deliver the most probable answer based on the available data and learned patterns. They do not directly consider user feedback or critical thinking. Therefore, current generative AI systems only provide information where the systems themselves can be highly certain of its accuracy. Consequently, useful but potentially unsafe answers may be withheld by current AI systems.The present invention is therefore based on the objective of optimizing the provision of information using artificial intelligence, particularly with regard to the variety of information provided.

[0007] This task is solved by a system for answering questions according to claim 1 and a method for answering questions according to claim 9.

[0008] The proposed question-answering system comprises a receiving unit for receiving a user's question and a processing unit for generating an answer to the user's question using artificial intelligence. The processing unit is further equipped to output the generated answer.

[0009] The question and answer can be entered and output in different ways, for example via various input / output devices that enable human-machine communication. These include, among others, text-based input (via keyboard or similar), voice input (via microphone, etc.), or voice output (via loudspeaker, etc.).

[0010] The artificial intelligence used can be any AI system, particularly one that represents "generative artificial intelligence." Such AI systems are those capable of generating text, images, and / or videos, as well as other data, using generative models. These systems utilize complex algorithms to learn patterns, relationships, and data from existing datasets and, based on this, generate new data or content that corresponds to the learned patterns but is not identical to them. Neural networks, in particular, can be used for this purpose.

[0011] As explained above, current AI systems provide information or answers that are highly likely to be correct. However, to account for the fact that users are more or less attentive and can, in principle, evaluate AI system responses based on their content, the proposed system uses AI to provide answers and considers a user's reaction to subsequent responses.

[0012] The processing unit is designed to output a response with a predefined confidence level. In this context, a confidence level represents the probability that the response is correct. A high confidence level indicates that the response is highly likely to be correct (e.g., between 90 and 100%), while a lower confidence level indicates that the response is less likely to be correct. The confidence level can be adjusted to allow for a more precise definition of the probability of correctness.

[0013] The user can view and rate the generated response. The receiving unit can receive feedback from the user, such as a correction or a question, and the processing unit can then adjust the predefined security score based on this feedback. If the receiving unit does not receive any user feedback, the processing unit continues with the predefined security score without adjustment.

[0014] After adjusting the security code, the processing unit can modify the output response and issue a customized answer. This can occur, for example, in response to user input requesting a customized response. Furthermore, the processing unit can use the customized security code for future responses.

[0015] This system increases the adaptability of the AI ​​used, as it allows for consideration of how a user reacts to the provided answers. This means that if the user is attentive and evaluates the answer, thus providing relevant feedback, the processing unit can react accordingly and adapt the answers, and therefore the information delivery, to this user interaction. Overall, this allows for an increase in the scope of information and an improvement in information quality without increasing the risk of incorrect answers.

[0016] The respective unit, for example, a receiving unit or a processing unit, can be implemented in hardware and / or software. In a hardware implementation, the respective unit can be a device or part of a device, for example, a computer, a microprocessor, or a control unit within a system or on a server, a host system, or similar. In a software implementation, the respective unit can be a computer program product, a function, a routine, part of program code, or an executable object.

[0017] According to one embodiment, the processing unit is configured to examine the user's feedback and determine a level of criticism, and the processing unit is further configured to adjust the predefined safety index based on the level of criticism.

[0018] Such a criticism level can indicate the extent to which a user reacts to a given response and can thus be an indicator of the user's attention. For example, according to one embodiment, the processing unit can be configured to examine user feedback regarding the number of queries and / or corrections to the given response and to adjust the criticism level based on this number. A high number of queries and / or corrections can, in particular, reflect that the user is attentive, and the criticism level can be correspondingly high. Few or no queries and / or corrections, on the other hand, can indicate that the user is inattentive or at least not interacting with the system, and the criticism level can be correspondingly low. Thus, the user's attention can be directly taken into account when adjusting the security metric.

[0019] According to another embodiment, the processing unit is configured to decrease the predefined security rating when the specific level of criticality exceeds a predefined threshold, and to increase the predefined security rating when the specific level of criticality falls below a predefined threshold. Thus, at a low level of criticality, i.e., a rather inattentive user or a user with few responses, the processing unit can increase the security rating. As a consequence, the processing unit can output responses based on this higher security rating, i.e., responses that are more likely to be correct. Conversely, at a high level of criticality, i.e., an attentive user or a user with many responses or interactions, the processing unit can decrease the security rating.As a result, the processing unit can output answers based on this lower confidence level, i.e., answers that are less likely to be correct.

[0020] According to another embodiment, the processing unit is configured to output a response with lower certainty for a lower security index than for a higher security index. As explained above, this assumes that an attentive user may accept responses with lower certainty because they will critically examine and evaluate the answers. This has the advantage of providing more comprehensive information in this case. Conversely, a higher security index allows for fewer, but more accurate and correct, responses to be output.

[0021] The system thus enables the responses output by the processing unit to be dynamically adapted based on the user's level of critical engagement. User behavior can be analyzed in real time, particularly since feedback on output responses can be processed directly.

[0022] According to a further embodiment, the processing unit is configured to identify portions of the response that correspond to the predefined or adjusted security level and portions of the response that have a security level below or above the predefined or adjusted security level. In particular, when the security level is low, the processing unit can identify portions of the response that correspond to a higher security level, and thus are more likely to be correct. In this way, a user does not have to check all parts of the response but can concentrate on the portions that are less likely to be correct.

[0023] According to another embodiment, the processing unit is designed to examine the user's feedback on several metrics and to optimize the predefined security indicator and the content of further responses based on the examined metrics.

[0024] As explained above, the number or frequency of queries and / or corrections can be used as a metric. Furthermore, the processing unit can identify the depth of the queries / corrections to understand whether they are basic or complex. Additionally, the diversity of the questions, queries, and / or corrections can be captured to determine whether the user is interacting at different levels.

[0025] According to another embodiment, the processing unit is configured to adjust the security rating based on each question-answer combination. The processing unit can thus continuously adjust the security rating to reflect the user's current level of attention. Furthermore, this enables feedback-based correction loops.

[0026] The processing unit can thus modify its response strategy based on the captured user feedback. In the case of critical user behavior, the processing unit can output more informative, but less certain, answers.

[0027] Feedback-based correction loops allow for the implementation of mechanisms to review and, if necessary, correct one's own answers, especially when the user identifies errors. This enables the processing unit to improve the quality of its answers while simultaneously increasing the level of information for the user, without significantly increasing the risk of misinformation. This can lead to greater user satisfaction and added value for critical users. This process can also be made transparent to encourage users to critically examine the answers. As mentioned previously, for example, parts of the answer can be marked as certain or uncertain.

[0028] In one example, a user asks the processing unit a question to get help generating code for statistical analysis. The processing unit responds with the safest answer it can currently provide, but this doesn't completely solve the user's problem. Next, the user might react critically and point out errors in the processing unit's answer. In this example, the user then asks another question about generating code. The processing unit notices the user's critical behavior, lowers its confidence score, and again provides code, but goes a step further by delivering code snippets that are potentially incorrect but could be a step in the right direction. The code in the output is indeed faulty, but the code snippets help the user quickly find the correct answer.The user now notes again what was wrong and what solution they found. The processing unit can now use this information not only to adjust the security indicator, but also to improve the generative AI in general by adapting the data that the AI ​​uses.

[0029] According to another aspect, a procedure for answering questions is proposed. The procedure comprises the following steps:

[0030] Receiving a question from a user,

[0031] Generating an answer to the user's question using artificial intelligence,

[0032] Outputting the answer,

[0033] Receiving feedback from the user on the issued response, where the response has a predefined security identifier, and

[0034] Adjusting the predefined safety rating based on the feedback. The embodiments and features described for the proposed system apply accordingly to the proposed method.

[0035] Furthermore, a computer program product is proposed which has program code designed to initiate the execution of the procedure described above on a computer.

[0036] A computer program product, such as a computer program tool, can be provided or delivered from a server on a network, for example, as a storage medium such as a memory card, USB stick, CD-ROM, DVD, or as a downloadable file. This can be done, for example, in a wireless communication network by transmitting the corresponding file containing the computer program product or tool.

[0037] Other possible implementations of the invention also include combinations of features or embodiments described previously or subsequently with regard to the exemplary embodiments, even if not explicitly mentioned. In such cases, the person skilled in the art will also add individual aspects as improvements or additions to the respective basic form of the invention.

[0038] Further advantages and advantageous embodiments are specified in the description, the drawings, and the claims. In particular, the combinations of features specified in the description and the drawings are purely exemplary, so that the features may also exist individually or in different combinations.

[0039] The invention will now be described in more detail with reference to exemplary embodiments illustrated in the drawings. These exemplary embodiments and the combinations shown in them are purely illustrative and do not define the scope of protection of the invention. The scope of protection is defined solely by the appended claims.

[0040] Figure 1 shows a schematic block diagram of a system for answering questions; and

[0041] Fig. 2: a schematic flowchart of a procedure for answering questions.

[0042] Fig. 1 shows a system 1 for answering questions F. A user can ask questions F which are answered by the system 1 using artificial intelligence, in particular generative artificial intelligence.

[0043] System 1 has a receiving unit 2 that receives a question F. A processing unit 4 can answer this question F using artificial intelligence and output a corresponding answer A.

[0044] When generating answer A, processing unit 4 considers a security index SK. This security index SK defines the probability that answer A is correct. A low security index SK corresponds to a lower probability of answer A being correct, while a higher security index SK corresponds to a higher probability of answer A being correct.

[0045] The user can provide feedback (FB) in response to answer A. This feedback (FB) is received by receiving unit 2. Processing unit 4 can then use the feedback (FB), e.g., a query regarding answer A or a correction to answer A, to adjust the security code SK and generate an adjusted security code SK'. This adjusted security code SK' can then be used when generating the next answer A.

[0046] If the user is attentive and makes many questions or corrections, processing unit 4 can decrease the security index SK, and the next answer may be output with a lower probability of being correct. Conversely, if the user is rather inattentive, processing unit 4 can increase the security index SK, thus outputting answers with a higher probability of being correct.

[0047] Fig. 2 shows a corresponding procedure for answering questions F. The procedure comprises the following steps.

[0048] In a first step S1, a question F from a user is received.

[0049] Subsequently, in step S2, an answer A to the user's question F is generated using artificial intelligence and output in step S3.

[0050] In step S4, a response FB from the user can be received in response to the output answer A.

[0051] The predefined security indicator SK can then be adjusted in step S5 based on the feedback FB.

[0052] As shown in Fig. 2, after step S5 the procedure can return to step S1 and receive a new question F, which is then answered with the adjusted security identifier SK'. Alternatively, after step S5 the procedure can also return to step S2 and answer the previously asked question F again with the adjusted security identifier SK' or adjust the already given answer A accordingly.

[0053] The following is a specific example of the method shown in Fig. 2, which is the one shown in Fig.

[0054] System 1 is used and described.

[0055] Step S1: The user uploads a photo of a traffic situation with various objects, including traffic lights. They are asked to identify all the traffic lights in the picture.

[0056] Step S2: The processing unit analyzes the image and identifies only the traffic lights it is very certain it has recognized. The processing unit marks the traffic lights it is very certain it has recognized. Step S3: This is output as a response.

[0057] Step S4: No user feedback is received.

[0058] Step S5: Since the processing unit does not receive any critical feedback from the user, it retains the security indicator.

[0059] In the next iteration of the procedure (same question), an answer is generated again in step S2 and output in step S3.

[0060] Step S4: The user notices missing traffic lights and provides appropriate feedback.

[0061] Step S5: The processing unit adjusts the security indicator SK.

[0062] In the next iteration, the adjusted (reduced) safety factor SK' will be used.

[0063] Step S2: With the reduced confidence level SK', the processing unit ventures to identify additional traffic lights where it is less certain. The processing unit marks traffic lights where it is less certain and traffic lights where it is very certain that it has correctly identified them.

[0064] Step S3: The corresponding answer A is output.

[0065] Step S4: The user checks the image and makes corrections if necessary – removing false positives and confirming correct identifications. This is received as feedback.

[0066] Step S5: The processing unit readjusts the safety indicator SK' based on the user's latest interactions. In doing so, the processing unit learns which type of traffic light recognition the user prefers and which error rates are acceptable. The processing unit now delivers results that better balance accuracy with the user's desired level of information. The safety indicator can be adjusted as often as needed, and the overall behavior of the processing unit can be optimized.

[0067] In summary, the system and procedure described here create the possibility of incorporating user feedback when generating answers. This allows answers to be tailored to a user's attention span and optimizes their informational content.

[0068] 1 system

[0069] 2 receiver units

[0070] 4 processing units

[0071] A response

[0072] Question

[0073] Facebook feedback

[0074] SK safety code

[0075] SK' adapted safety indicator S1-S5 process steps

Claims

System for answering questions Claims:

1. System (1) for answering questions (F), with a receiving unit (2) for receiving a question (F) from a user, and a processing unit (4) for generating an answer (A) to the question (F) of the user using artificial intelligence and for outputting the answer (A), characterized in that that the answer (A) has a predefined security identifier (SK), that the receiving unit (2) is designed to receive feedback (FB) from the user on the output response (A), and that the processing unit (4) is trained to adjust the predefined security key number (SK) based on the feedback (FB).

2. System according to claim 1, wherein the processing unit (4) is configured to examine the user's feedback (FB) and determine a level of criticism, wherein the processing unit (4) is further configured to adjust the predefined security index (SK) based on the level of criticism.

3. System according to claim 2, wherein the processing unit (4) is configured to examine the user's feedback (FB) regarding the number of queries and / or corrections and to adjust the level of criticism based on the number of queries and / or corrections.

4. System according to claim 2 or 3, wherein the processing unit (4) is configured to decrease the predefined security index (SK) when the specified level of criticism is above a predefined threshold, and to increase the predefined security index (SK) when the specified level of criticism is below a predefined threshold.

5. System according to any one of the preceding claims, wherein the processing unit (4) is configured to output a response (A) with a lower level of certainty at a lower security index (SK) than at a higher security index (SK).

6. System according to one of the preceding claims, wherein the processing unit (4) is configured to identify portions of the response (A) that correspond to the predefined or adapted security index (SK, SK') and portions of the response that have a security index (SK) that is below or above the predefined or adapted security index (SK, SK').

7. System according to one of the preceding claims, wherein the processing unit (4) is configured to examine the user's feedback (FB) for several metrics and to optimize the predefined security indicator (SK) and the content of further responses (A) based on the examined metrics.

8. System according to any of the preceding claims, wherein the processing unit (4) is configured to adjust the security identifier (SK) based on each question-answer combination.

9. Procedure for answering questions (F), comprising the steps: Receiving (S1) a question (F) from a user, Generating (S2) an answer (A) to the user's question (F) using artificial intelligence, and Output (S3) of the answer (A), the procedure being characterized by the following further steps: Receiving (S4) a response (FB) from the user to the issued answer (A), where the answer (A) has a predefined security number (SK), and Adjusting (S5) the predefined security identifier (SK) based on the feedback (FB).

10. Computer program product comprising program code configured to execute the method of claim 9 on a computer.

Images