Device and method
The apparatus and method simplify the estimation and correction of position bias in generative AI for multiple-choice questions by modifying options to the same content and using a statistical model, enhancing accuracy and reducing costs.
Patent Information
- Application Number
- PCT/JP2025/017213
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-05-12
- Publication Date
- 2026-11-19
AI Technical Summary
Existing methods for estimating position bias in generative AI for multiple-choice questions are cumbersome and inefficient, often requiring complex calculations and increased API usage costs.
An apparatus and method that acquires multiple-choice questions, modifies all options to the same content, estimates position bias based on the probability of appearance of these modified options, and corrects the bias using a statistical model trained on feature information.
Enables easier, more accurate, and cost-effective estimation and removal of position bias in generative AI responses, improving response accuracy and reducing API usage costs.
Smart Images

Figure JP2025017213_19112026_PF_FP_ABST
Abstract
Description
Apparatus and Method
[0001] One aspect of the present disclosure relates to an apparatus and method for estimating the position bias of generative AI for multiple-choice questions.
[0002] The following Patent Document 1 discloses an information processing apparatus for estimating position bias.
[0003] Japanese Unexamined Patent Application Publication No. 2025-21419
[0004] The information processing apparatus estimates the position bias by executing the following complex processing. That is, the apparatus acquires placement probabilities representing the probabilities that each of n (n being a natural number of 2 or more) items is placed at k (k being a natural number of 2 or more) positions, converts the n items into m (m being a natural number of 2 or more) embedding vectors representing abstract representations of features of the n items, calculates assignment probabilities representing the probabilities of assignment from the n items to the m embedding vectors, and estimates, based on the assignment probabilities, a position bias representing the probability that a user recognizes each of the k positions where the m embedding vectors are placed.
[0005] Accordingly, there is a demand for more easily estimating the position bias.
[0006] An apparatus according to one aspect of the present disclosure includes an acquisition unit that acquires a multiple-choice question, and an estimation unit that estimates the position bias of the generative AI for the question, based on appearance probabilities of the options of a modified question output by the generative AI for the modified question in which all options of the question acquired by the acquisition unit are modified to have the same content.
[0007] In such an aspect, the position bias of the generative AI for the multiple-choice question is estimated based on the appearance probabilities of the options of the modified question output by the generative AI for the modified question in which all options of the multiple-choice question are modified to have the same content. That is, the position bias can be estimated more easily.
[0008] According to one aspect of the present disclosure, the position bias can be estimated more easily.
[0009] This figure shows an example of the functional configuration of the device according to the embodiment. This figure shows an example of a prompt for a multiple-choice question. This figure shows an example of the relationship between the device according to the embodiment and the generating AI. This figure shows a modified version of the relationship in Figure 3. This figure shows an example of a prompt for a modified question. This figure shows an example of a table of occurrence probabilities with position bias for a modified question. This figure shows an example of a table of occurrence probabilities without position bias for a modified question. This figure shows an example of a table for position bias correction. This figure shows an example of a table of occurrence probabilities with position bias for a multiple-choice question. This figure shows an example of a table of estimated occurrence probabilities without position bias for a multiple-choice question. This figure shows an example of a table of training data for the estimation model. This figure shows an example of a table of input data to the estimation model. This flowchart shows an example of a correction process performed by the device according to the embodiment. This flowchart shows an example of a training process performed by the device according to the embodiment. This figure shows an example of the hardware configuration of the computer used in the device according to the embodiment.
[0010] The embodiments of this disclosure will be described in detail below with reference to the drawings. In the description of the drawings, the same elements will be denoted by the same reference numerals, and redundant descriptions will be omitted. Furthermore, the embodiments of this disclosure described below are specific examples of the present invention and are not limited to these embodiments unless otherwise stated to limit the present invention.
[0011] Figure 1 shows an example of the functional configuration of the apparatus 1 according to the embodiment. Apparatus 1 is a computer device (server) that estimates the position bias of the generating AI for multiple-choice questions and removes the position bias by correcting it.
[0012] Multiple-choice questions are also called multiple-choice questions. Multiple-choice or multiple-choice methods involve providing multiple options (or items; the same applies hereinafter) and having the user select the appropriate option from among them, or a method thereof. In the embodiments, it is assumed that the user will select one option from multiple options, but this is not limited to this, and one or more options may be selected from multiple options. A multiple-choice question consists, for example, of a question sentence and multiple options. A multiple-choice question may consist of multiple options (only). A multiple-choice question may be text, or of course, a prompt composed of text. In the embodiments, a multiple-choice question is assumed to be a prompt, but this is not limited to this.
[0013] A prompt is information indicating instructions or questions to be entered by the user in an interactive system such as a dialogue with a generating AI or a command-line interface (CLI). For example, a prompt may express, in text form, the command to be executed by the generating AI, the task to be performed by the generating AI, the background / context that the generating AI should consider (e.g., role, conditions), the question that the generating AI should answer, and the output format of the response information from the generating AI. In addition, input information that is the target of the command / task to be executed by the generating AI may be attached to the prompt. Such input information includes data files with filenames containing predetermined extensions, such as text data, image data, application-related data, audio data, video data, and still image data. Application-related data refers to data such as document data, table data, and graph data that can be processed by a default application program.
[0014] Figure 2 shows an example of a prompt for a multiple-choice question. The prompt example shown in Figure 2 consists of the question "You are a Japanese man in your 20s. Which sushi topping is your favorite?" and three options: 1. "Tuna", 2. "Salmon", and 3. "Salmon roe". In this embodiment, "Tuna", "Salmon", or "Salmon roe" (the actual options), excluding the part that identifies the option (the number and period at the beginning), are simply referred to as "Options".
[0015] Generative AI is a technology that, in response to a prompt (or text; the same applies hereinafter) containing input information, generates content according to one or a combination of the instructions, context, questions, and output formats indicated by the prompt, and returns that content as response information. The prompt can also include input information, in which case the generative AI generates response information targeting the input information. The generative AI may be a conversational AI that includes, for example, a Large Language Model (LLM) and a User Interface (UI) for interaction with the user, enabling text chat or voice chat with the user. Examples of such generative AI include tsuzumi®, ChatGPT, GPT®-3.5, GPT-4V, and PaLM2.
[0016] The generated AI may be stored in another device connected to device 1 via a network, and configured to allow for information exchange. Figure 3 shows an example of the relationship between device 1 and the generated AI. In the relationship shown in Figure 3, the generated AI, LLM20, is stored in an external server 2 connected to device 1 via a network. When needed, device 1 has LLM20 perform processing via the network, receives the execution results, and uses them.
[0017] The generated AI may be stored within the device 1. Figure 4 shows a modified example of the relationship shown in Figure 3. In the modified example shown in Figure 4, the generated AI, LLM20, is stored within the device 1 (for example, in the storage unit 10). When necessary, the device 1 executes processing on the LLM20 stored within the device 1 and uses the execution result.
[0018] Position bias is the tendency for an option to be selected more often when it is located in a particular position compared to options located elsewhere. In a task where a generative AI (LLM20) is asked multiple-choice questions, position bias exists, making it easier to select a specific numbered option (for example, option 1, "Tuna," in Figure 2). For example, when asked a multiple-choice question with two options, such as "You are Japanese. Do you eat bread or rice for breakfast? 1. Bread 2. Rice," there is a tendency for responses to gravitate towards option "1."
[0019] Device 1 may also have the function of RAG (Retrieval-Augmented Generation).
[0020] Returning to Figure 1, as shown in Figure 1, the device 1 is composed of a storage unit 10, an acquisition unit 11 (acquisition unit), a modification unit 12 (modification unit), an estimation unit 13 (estimation unit), a correction unit 14 (correction unit), and a generation unit 15 (generation unit).
[0021] Each functional block of Device 1 is intended to function within Device 1, but is not limited to this. For example, some of the functional blocks of Device 1 may function within a computer device separate from Device 1, connected to Device 1 via a network, while appropriately sending and receiving information with Device 1. Furthermore, some functional blocks of Device 1 may be omitted, multiple functional blocks may be integrated into a single functional block, or a single functional block may be broken down into multiple functional blocks.
[0022] The following describes the functions of the device 1 shown in Figure 1.
[0023] The storage unit 10 stores arbitrary information used in calculations in the device 1, as well as the results of calculations in the device 1. The information stored in the storage unit 10 may be referenced as appropriate by each function of the device 1. The information stored in the storage unit 10 may be pre-set information, or it may be information acquired at any time by the acquisition unit 11 described later.
[0024] The acquisition unit 11 acquires arbitrary information. The acquisition unit 11 may acquire arbitrary information used in calculations in the device 1 from other functional blocks of the device 1, including the storage unit 10, or it may acquire (receive) it from other devices via a network or the like. The acquisition unit 11 may acquire arbitrary information in advance prior to any calculations in the device 1, or it may acquire the information when it becomes necessary, in accordance with the requests of each functional block of the device 1. The acquisition unit 11 may store the acquired information in the storage unit 10, or it may output it to other functional blocks of the device 1.
[0025] The acquisition unit 11 acquires multiple-choice questions. In this embodiment, multiple-choice questions are simply referred to as "questions" as appropriate. The acquisition unit 11 may acquire questions from the user via the input device 1005 described later, or from other devices via the communication device 1004, or it may acquire questions stored by the storage unit 10.
[0026] The acquisition unit 11 may store the acquired questions in the storage unit 10, or it may output them to other functional blocks of the device 1.
[0027] The modification unit 12 generates a modified question in which all the answer choices of the question acquired by the acquisition unit 11 are changed to the same content. Changing all the answer choices to the same content means, for example, changing all the answer choices to the same content as the nth answer choice (where n is a natural number starting from 1, and the maximum number is the number of choices), or changing all the answer choices to arbitrary content (including content different from each answer choice), etc.
[0028] Figure 5 shows an example of a modified question prompt. More specifically, the modified question prompt shown in Figure 5 is generated by changing all the options in the question prompt example shown in Figure 2 to the first option, "tuna" (by the modification unit 12). The modified question prompt shown in Figure 5 consists of the same question text (without modification) as the question prompt example shown in Figure 2: "You are a Japanese man in your 20s. Which sushi topping do you like best?" and three options: "1. Tuna", "2. Tuna", and "3. Tuna". Thus, in the modified question prompt example shown in Figure 5, the second and third options are changed to "tuna" compared to the question prompt example shown in Figure 2. In other words, all options are changed to the same "tuna".
[0029] The modification unit 12 may store the generated modification questions in the storage unit 10, output them to other functional blocks of the device 1, output (transmit) them to other devices via the communication device 1004 described later, or output (display) them to the output device 1006 described later.
[0030] The estimation unit 13 estimates the position bias of the generating AI towards a question based on the probability of appearance of the options for the modified question output by the generating AI for a modified question in which all options of the question acquired by the acquisition unit 11 have been changed to the same content. The estimation unit 13 may also estimate the position bias of the generating AI towards a question based on the probability of appearance of the options for the modified question output by the generating AI for a modified question generated by the modification unit 12.
[0031] The modification questions used by the estimation unit 13 may be generated by the modification unit 12, generated by any functional block of the device 1 including the estimation unit 13, or generated by another device. The probability of the options for the modification questions used by the estimation unit 13 may be output by the generating AI when an arbitrary functional block of the device 1 including the estimation unit 13 inputs a modification question to the generating AI, or output by the generating AI when another device inputs a modification question to the generating AI.
[0032] The estimation unit 13 may estimate based on the difference between the probability of appearance of the options for the modified question output by the generating AI in response to the modified question and the probability of appearance if all options for the modified question had the same probability of appearance. The estimation unit 13 may also estimate the position bias of the generating AI with respect to one option of the question corresponding to the above one option (for example, the second option of the modified question) (for example, the second option of the question; the second options correspond to each other) based on the difference between the probability of appearance of one option of the modified question output by the generating AI in response to the modified question and the probability of appearance if all options for the modified question had the same probability of appearance.
[0033] The generating AI may be controlled to select one of the options in a multiple-choice question. For example, if a question consists of three options, the generating AI may be controlled to select only one of the three options (i.e., it can only select one of the three options). This control may be implemented by configuring the generating AI. If the generating AI is OpenAI®, this control may be implemented in the API (Application Programming Interface) "Structured Outputs," etc.
[0034] The estimation unit 13 may store the estimated position bias or information based on the position bias in the storage unit 10, output it to another functional block of the device 1, output (transmit) it to another device via the communication device 1004 described later, or output (display) it to the output device 1006 described later.
[0035] The correction unit 14 corrects the probability of the options for a question output by the generating AI based on the position bias estimated by the estimation unit 13. For example, the correction unit 14 corrects the probability of the options for a question output by the generating AI by adding or subtracting the position bias estimated by the estimation unit 13 (or performing any other arbitrary operation). The correction by the correction unit 14 can remove the position bias (debias, remove bias).
[0036] The correction unit 14 may store the position vice removal result in the storage unit 10, output it to another functional block of the device 1, output (transmit) it to another device via the communication device 1004 described later, or output (display) it to the output device 1006 described later.
[0037] Specific examples of the processing performed by the estimation unit 13 and the correction unit 14 will be explained with reference to Figures 6 to 10.
[0038] Figure 6 shows an example table of position-biased occurrence probabilities for modified questions. In the example table shown in Figure 6, the numbers in the "Answer" column indicate the nth number (nth n) included in the options of the modified question, and the percentages in the "Occurrence Probability" column indicate the occurrence probability output by the generating AI for that option. The occurrence probabilities shown in Figures 6 to 10 were confirmed using OpenAI® API's "Logprobs" or "top_logprobs," but are not limited to these.
[0039] The example table shown in Figure 6 is an example of a table showing the probability of appearance of the options for the modified question output by the generating AI in response to the example prompt for the modified question shown in Figure 5. The example table shown in Figure 6 shows that the generating AI output an appearance probability of "44%" for option "1. "Tuna"" (the "Answer" column refers to the record containing the number "1" in that option), that the generating AI output an appearance probability of "32%" for option "2. "Tuna"" (the "Answer" column refers to the record containing the number "2" in that option), and that the generating AI output an appearance probability of "24%" for option "3. "Tuna"" (the "Answer" column refers to the record containing the number "3" in that option).
[0040] Figure 7 shows an example table of the probability of occurrence without position bias for a modified question. More specifically, the example table in Figure 7 shows the probability of occurrence without position bias for the modified question prompt example shown in Figure 5. Without position bias, the content of all options is the same, so the probability of occurrence for each option is equal. In the modified question prompt example shown in Figure 5, there are three options, so the probability of occurrence for each option is "33.33%" (decimal places beyond the third are omitted; the same applies below), which is "100%" divided by the number of options, "3". This "33.33%" is the probability of occurrence if each option in the modified question shown in Figure 5 had the same probability of occurrence. The example table in Figure 7 shows that the probability of occurrence without position bias for each option in the modified question prompt example shown in Figure 5 is "33.33%".
[0041] Figure 8 shows an example of a table for correcting position bias. The example table in Figure 8 shows the position bias of the generating AI for each question, as estimated by the estimation unit 13. The example table in Figure 8 shows the difference between the probability of occurrence in the example table of position bias for the modified question shown in Figure 6 and the probability of occurrence in the example table of position bias for the modified question shown in Figure 7.
[0042] For example, the amount (or value) of the position bias of the record whose "Answer" column is "1" in the example table shown in FIG. 8 is "+10.66%", which is obtained by subtracting the occurrence probability "33.33%" of the record whose "Answer" column is "1" in the example table shown in FIG. 7 from the occurrence probability "44%" of the record whose "Answer" column is "1" in the example table shown in FIG. 6. Similarly, the amount (or value) of the position bias of the record whose "Answer" column is "2" in the example table shown in FIG. 8 is "-1.33%", which is obtained by subtracting the occurrence probability "33.33%" of the record whose "Answer" column is "1" in the example table shown in FIG. 7 from the occurrence probability "32%" of the record whose "Answer" column is "2" in the example table shown in FIG. 6. Similarly, the amount (or value) of the position bias of the record whose "Answer" column is "3" in the example table shown in FIG. 8 is "-9.33%", which is obtained by subtracting the occurrence probability "33.33%" of the record whose "Answer" column is "1" in the example table shown in FIG. 7 from the occurrence probability "24%" of the record whose "Answer" column is "3" in the example table shown in FIG. 6.
[0043] FIG. 9 is a diagram illustrating an example table of occurrence probabilities with position bias for multiple-choice questions. In the example table shown in FIG. 9, the numbers in the "Answer" column indicate the numbers (the n-th n) included in the options of the question, and the percentages in the "Occurrence Probability" column indicate the occurrence probabilities output by generative AI for the corresponding options.
[0044] The example table shown in FIG. 9 is an example table of occurrence probabilities of the options of the question output by generative AI in response to the example question prompt shown in FIG. 2. The example table shown in FIG. 9 indicates that generative AI output an occurrence probability of "56%" for the question option "1. 'Tuna'" (see the record whose "Answer" column is the number "1" included in the corresponding option), indicates that generative AI output an occurrence probability of "38%" for the question option "2. 'Salmon'" (see the record whose "Answer" column is the number "2" included in the corresponding option), and indicates that generative AI output an occurrence probability of "6%" for the modified question option "3. 'Salmon roe'" (see the record whose "Answer" column is the number "3" included in the corresponding option).
[0045] FIG. 10 is a diagram illustrating an example table of estimated occurrence probabilities without position bias for multiple-choice questions. Although the example table shown in FIG. 10 presents calculation formulas for clarity, the calculation results of the calculation formulas may be presented instead of the calculation formulas. For example, the "occurrence probability" of the record whose "answer" column is "1" may be set to "45.33%", which is the calculation result of the calculation formula "56-10.66%", the "occurrence probability" of the record whose "answer" column is "2" may be set to "39.33%", which is the calculation result of the calculation formula "38+1.33%", and the "occurrence probability" of the record whose "answer" column is "3" may be set to "15.33%", which is the calculation result of the calculation formula "6+9.33%".
[0046] The example table shown in FIG. 10 indicates the occurrence probabilities of the options for the question output by generative AI in response to the question, which have been corrected by the correction unit 14. The example table shown in FIG. 10 is obtained by subtracting the position bias of the example table of the position bias correction table shown in FIG. 8 from the occurrence probabilities of the example table of occurrence probabilities for the question shown in FIG. 9. Such correction may be implemented using, for example, "logit_bias" of the API of OpenAI (registered trademark).
[0047] As described above, when generative AI is simply caused to output occurrence probabilities for the question shown in FIG. 2, the result is the occurrence probabilities with position bias as shown in FIG. 9. Therefore, by executing the above-described functions of the device 1, it is possible to estimate the occurrence probabilities without position bias as shown in FIG. 10.
[0048] The specific examples of the processing of the estimation unit 13 and the correction unit 14 have been described above.
[0049] The generation unit 15 generates an estimation model for multiple-choice questions by learning training data that includes feature information relating to one or more features of a question and the position bias of the generating AI for that question estimated by the estimation unit 13. The generation unit 15 may learn using the feature information included in the training data as explanatory variables and the position bias included in the training data as the objective variable. The estimation unit 13 may estimate the position bias of the generating AI for that question by inputting the feature information of the question acquired by the acquisition unit 11 into the estimation model generated by the generation unit 15.
[0050] Specific examples of question characteristics include the question embedding (embedding of the question text), the question category (type, category, classification, genre), the number of question tokens (the basic unit used by LLM when processing text) for the question (or the question text contained within it), and the number of answer choices included in the question.
[0051] Figure 11 shows an example of a training data table (position bias database) for an estimation model. In the example training data table shown in Figure 11, the question text (content) included in a question is associated with the question embedding, the question category, the number of question tokens, and the position bias (in "%") of the generating AI for each option in that question. The question embedding, question category, and number of question tokens are calculated based on the question using existing technology or general methods. Each position bias may be estimated by the estimation unit 13, or it may be estimated or set by other means. The generation unit 15 may generate an estimation model by learning the position bias of the generating AI for each option with the training data shown in Figure 11 as explanatory variables and the position bias of the generating AI for each option as the target variable (learning the value of the position bias).
[0052] Figure 12 shows an example of a table of input data for the estimation model. In the example input data table shown in Figure 12, the question text (content) of the question for which position bias is to be estimated corresponds to the question embedding, the question category, and the number of question tokens for that question. The columns corresponding to each option of the question (three columns, as three options are assumed) are blank (not entered). As described above, the question embedding, question category, and number of question tokens are calculated using existing technology or general methods based on the question for which position bias is to be estimated. The estimation unit 13 inputs the question embedding, question category, and number of question tokens (these are characteristic information of the question for which position bias is to be estimated) and the three blanks from the example input data table shown in Figure 12 into the estimation model trained based on the training data shown in Figure 11. In response, the estimation model inputs the estimated position bias of the generating AI for each option of the question for which position bias is to be estimated into the three blanks. The estimation unit 13 estimates the position bias as the position bias of the generating AI for each option in the question for which the position bias is to be estimated (estimates the position bias for the question).
[0053] The estimation described above using Figures 11 and 12 can also be considered statistical estimation using a position bias database. While Figures 11 and 12 are based on learning and estimation for questions with three options, the process is not limited to three options; similar learning and estimation can be performed for any number of options.
[0054] The generation unit 15 may store the generated estimation model in the storage unit 10, output it to another functional block of the device 1, or output (transmit) it to another device via the communication device 1004 described later.
[0055] Next, an example of the correction process (including estimation process) performed by device 1 will be explained with reference to Figure 13. Figure 13 is a flowchart showing an example of the correction process performed by device 1.
[0056] First, the acquisition unit 11 acquires a multiple-choice question (step S1, acquisition step). Next, the modification unit 12 generates a modified question in which all the options of the question acquired in step S1 are changed to the same content (step S2, modification step). Next, the estimation unit 13 estimates the position bias of the generating AI for the question acquired in step S1 based on the probability of appearance of the options of the modified question output by the generating AI for the modified question generated in step S2 (step S3, estimation step). Next, the correction unit 14 corrects the probability of appearance of the options of the question output by the generating AI for the question acquired in step S1 based on the position bias estimated in step S3 (step S4, correction step).
[0057] Step S2 may be omitted, and in step S3, the estimation unit 13 may estimate the position bias of the generating AI for the question acquired in step S1 based on the probability of appearance of the options for the modified question output by the generating AI for a modified question in which all options of the question acquired in step S1 have been changed to the same content. Step S4 may also be omitted. Furthermore, regarding the probability of appearance of the options for the question acquired in step S1, which is used in step S4, the output of the probability of appearance by the generating AI may be performed at any time (but before step S4), for example, before step S1. Furthermore, in step S3, the estimation unit 13 may output the estimated position bias or information based on the position bias. Furthermore, in step S4, the correction unit 14 may output the correction result or information based on the correction result.
[0058] Next, an example of the learning process performed by device 1 will be explained with reference to Figure 14. Figure 14 is a flowchart showing an example of the process performed by device 1.
[0059] First, the generation unit 15 collects training data including feature information relating to one or more features of a multiple-choice question and the position bias of the generating AI for that question estimated by the estimation unit 13 (step S10, collection step). Next, the generation unit 15 learns (machine learning) the training data collected in step S10 (step S11, learning step). Next, the generation unit 15 generates an estimation model that estimates the position bias of the generating AI for a multiple-choice question based on the learning in step S11 (step S12, generation step).
[0060] Furthermore, the position bias included in the training data collected in step S10 is not limited to that estimated by the estimation unit 13, but may be estimated by any other means or may be a preset bias. Also, in step S12, the generation unit 15 may output the generated estimation model or information based on the estimation model.
[0061] Next, the effects and advantages of the apparatus 1 according to this embodiment will be described.
[0062] The device 1 comprises an acquisition unit 11 that acquires multiple-choice questions, and an estimation unit 13 that estimates the position bias of the generating AI to a question based on the probability of appearance of the options in a modified question output by the generating AI for a modified question in which all the options of the acquired question have been changed to the same content. With this configuration, the position bias of the generating AI to a question is estimated based on the probability of appearance of the options in a modified question output by the generating AI for a modified question in which all the options of a multiple-choice question have been changed to the same content. In other words, the position bias can be estimated more easily. For example, the position bias can be easily estimated simply by inputting a modified question in which all the options of a multiple-choice question have been changed to the same content into the generating AI.
[0063] The estimation unit 13 may estimate (the position bias of the generating AI for the question acquired by the acquisition unit 11) based on the difference between the probability of appearance of (the options for the modified question output by the generating AI in response to the modified question) and the probability of appearance if all options for the modified question had the same probability of appearance. This configuration allows for more accurate and reliable estimation.
[0064] The estimation unit 13 may estimate the position bias of the generating AI for one of the options in the question (acquired by the acquisition unit 11) corresponding to one of the options in the modified question, based on the difference between the probability of one of the options in the modified question output by the generating AI in response to the modified question and the probability of all options in the modified question having the same probability of appearance. This configuration makes it possible to estimate the position bias of the generating AI for one of the options in the question more accurately and reliably.
[0065] The generating AI may be controlled to select one of the options for a multiple-choice question. This configuration allows for more accurate estimation of the probability of the options for a modified question output by the generating AI, and therefore more accurate estimation of position bias.
[0066] The device 1 may further include a correction unit 14 that corrects the probability of appearance of the answer choices for a question output by the generating AI in response to the question, based on the position bias estimated by the estimation unit 13. With this configuration, for example, the position bias can be removed from the probability of appearance of the answer choices for a question output by the generating AI in response to the question.
[0067] The device 1 may further include a modification unit 12 that generates a modified question in which all the options of the question acquired by the acquisition unit 11 have been changed to the same content, and the estimation unit 13 may estimate the position bias of the generating AI with respect to the question based on the probability of appearance of the options of the modified question output by the generating AI for the modified question generated by the modification unit 12. With this configuration, modified questions can be generated more reliably, and therefore the position bias can be estimated more reliably.
[0068] The device 1 may further include a generation unit 15 that generates an estimation model for multiple-choice questions by learning training data which includes feature information relating to one or more features of a question and the position bias of the generating AI for that question estimated by the estimation unit 13. With this configuration, for example, when feature information of a multiple-choice question is input, an estimation model that outputs the position bias of the generating AI for that question can be generated. This makes it easier to estimate the position bias.
[0069] The generation unit 15 may learn using the feature information contained in the training data as explanatory variables and the position bias contained in the training data as the target variable. With this configuration, for example, when feature information of a multiple-choice question is input, an estimation model can be generated that outputs the position bias of the generating AI for that question. This makes it easier to estimate the position bias.
[0070] The estimation unit 13 may estimate the position bias of the generating AI for a given question by inputting the characteristic information of the question acquired by the acquisition unit 11 into the estimation model generated by the generation unit 15. This configuration makes it possible to more reliably estimate the position bias of the generating AI for multiple-choice questions.
[0071] A challenge exists in tasks where LLMs are given multiple-choice questions, where position bias exists, making it easier to select a specific numbered option. Therefore, accurately evaluating the output accuracy of LLMs is currently difficult. A current solution involves "reversing the order of the options, asking multiple questions, and taking the representative value (mean or mode)," but this increases API usage costs and execution time (for n options, n! (n factorial) questions are required).
[0072] According to Device 1, for a multiple-choice question (multiple-choice question) from which position bias is to be removed, a question is prepared in which all the options are the same, and the probability of occurrence (output probability) of each option is calculated to estimate the amount of position bias for each option. Then, the estimated bias is corrected for the probability of occurrence of the original question. Device 1 also provides a method for statistically estimating the amount of position bias. Specifically, the embedding of the question text, the category of the question, and the number of tokens in the question text are used as explanatory variables, and the value of the position bias for each option is used as the dependent variable to construct a machine learning predictor.
[0073] Device 1 removes position bias when LLMs are given multiple-choice questions. Device 1 provides a method for removing LLM position bias from multiple-choice questions. According to Device 1, position bias can be easily removed with API usage fees equivalent to two questions (1st time: outputting the probability of occurrence with position bias initially; 2nd time: having the LLM respond to a modified question with all options the same), thereby improving response accuracy. Furthermore, while Device 1's method for statistically estimating the amount of position bias incurs an initial cost of API usage fees when preparing training data (dataset), after training the estimation model, debiasing becomes possible with API usage fees equivalent to one question.
[0074] The apparatus 1 of this disclosure may have the following configuration.
[0075] [1] An apparatus comprising: an acquisition unit that acquires multiple-choice questions; and an estimation unit that estimates the position bias of the generating AI with respect to the questions based on the probability of appearance of the options of the modified questions output by the generating AI in response to a modified question in which all the options of the questions acquired by the acquisition unit have been changed to the same content.
[0076] [2] The apparatus according to [1], wherein the estimation unit estimates based on the difference between the probability of occurrence and the probability of occurrence if all of the options in the modified question had the same probability of occurrence.
[0077] [3] The apparatus according to [1] or [2], wherein the estimation unit estimates the position bias of the generating AI for one of the options in the question corresponding to one of the options, based on the difference between the probability of appearance of one of the options in the modified question output by the generating AI in response to the modified question and the probability of appearance of all of the options in the modified question having the same probability of appearance.
[0078] [4] The apparatus according to any one of [1] to [3], wherein the generating AI is controlled to select one of the options for a multiple-choice question.
[0079] [5] The apparatus according to any one of [1] to [4], further comprising a correction unit that corrects the probability of appearance of the options for the question output by the generating AI in response to the question, based on the position bias estimated by the estimation unit.
[0080] [6] The apparatus according to any one of [1] to [5], further comprising a modification unit that generates a modified question in which all the options of the question acquired by the acquisition unit have been changed to the same content, and the estimation unit estimates the position bias of the generating AI with respect to the question based on the probability of appearance of the options of the modified question output by the generating AI with respect to the modified question generated by the modification unit.
[0081] [7] The apparatus according to any one of [1] to [6], further comprising a generation unit that generates an estimation model for estimating the position bias of the generation AI for a multiple-choice question by learning training data including feature information relating to one or more features of the question and the position bias of the generation AI for the question estimated by the estimation unit.
[0082] [8] The apparatus according to [7], wherein the generation unit learns the feature information contained in the training data as an explanatory variable and the position bias contained in the training data as an objective variable.
[0083] [9] The apparatus according to [7] or [8], wherein the estimation unit estimates the position bias of the generating AI for the question by inputting the characteristic information of the question acquired by the acquisition unit into the estimation model generated by the generation unit.
[0084] The block diagrams used in the description of the above embodiments show functional units. These functional blocks (components) are realized by any combination of at least one of hardware and software. Furthermore, the method of realizing each functional block is not particularly limited. That is, each functional block may be realized using one device that is physically or logically coupled, or it may be realized using two or more physically or logically separated devices that are directly or indirectly connected (for example, using wired or wireless connections). A functional block may also be realized by combining software with the one or more of the above devices.
[0085] Functions include, but are not limited to, judgment, decision, determination, calculation, calculation, processing, derivation, investigation, exploration, confirmation, reception, transmission, output, access, resolution, selection, selection, establishment, comparison, assumption, expectation, assumption, broadcasting, notifying, communicating, forwarding, configuring, reconfiguring, allocating (mapping), and assigning. For example, a functional block (configuration part) that enables transmission is called a transmitting unit or transmitter. In all cases, as mentioned above, the method of implementation is not particularly limited.
[0086] For example, Apparatus 1 in one embodiment of the present disclosure may function as a computer that processes the estimation method, correction method, or learning method of the present disclosure. Figure 15 is a diagram showing an example of the hardware configuration of Apparatus 1 according to one embodiment of the present disclosure. Physically, Apparatus 1 may be configured as a computer device including a processor 1001, memory 1002, storage 1003, communication device 1004, input device 1005, output device 1006, bus 1007, etc.
[0087] In the following explanation, the term "device" can be replaced with "circuit," "device," "unit," etc. The hardware configuration of device 1 may include one or more of the devices shown in the figure, or it may be configured to omit some of the devices.
[0088] Each function in device 1 is realized by loading predetermined software (programs) onto hardware such as the processor 1001 and memory 1002, which allows the processor 1001 to perform calculations, control communication by the communication device 1004, and control at least one of data reading and writing in the memory 1002 and storage 1003.
[0089] The processor 1001 controls the entire computer, for example, by running an operating system. The processor 1001 may be composed of a central processing unit (CPU) that includes interfaces with peripheral devices, control devices, arithmetic units, registers, etc. For example, the acquisition unit 11, modification unit 12, estimation unit 13, correction unit 14, and generation unit 15 described above may be implemented by the processor 1001.
[0090] Furthermore, the processor 1001 reads programs (program code), software modules, data, etc., from at least one of the storage 1003 and the communication device 1004 into the memory 1002 and executes various processes accordingly. The program used is one that causes the computer to execute at least a part of the operations described in the above embodiment. For example, the acquisition unit 11, the modification unit 12, the estimation unit 13, the correction unit 14, and the generation unit 15 may be implemented by a control program stored in the memory 1002 and running on the processor 1001, and other functional blocks may be implemented similarly. The above-described various processes have been explained as being executed by one processor 1001, but they may be executed simultaneously or sequentially by two or more processors 1001. The processor 1001 may be implemented by one or more chips. The program may also be transmitted from a network via a telecommunications line.
[0091] The memory 1002 is a computer-readable recording medium and may consist of at least one of the following: ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), RAM (Random Access Memory), etc. The memory 1002 may also be called a register, cache, main memory, etc. The memory 1002 can store executable programs (program code), software modules, etc., for carrying out a wireless communication method according to one embodiment of the present disclosure.
[0092] The storage 1003 is a computer-readable recording medium and may consist of at least one of the following: an optical disc such as a CD-ROM (Compact Disc ROM), a hard disk drive, a flexible disk, a magneto-optical disk (e.g., Compact Disc, Digital Multipurpose Disc, Blu-ray® Disc), a smart card, flash memory (e.g., a card, stick, key drive), a floppy® disk, a magnetic strip, etc. The storage 1003 may also be called an auxiliary storage device. The above-mentioned storage medium may be, for example, a database, server, or other suitable medium including at least one of memory 1002 and storage 1003.
[0093] The communication device 1004 is hardware (transceiver / receiver device) for communicating between computers via at least one of a wired network and a wireless network, and is also referred to as a network device, network controller, network card, communication module, etc. The communication device 1004 may be configured to include, for example, a high-frequency switch, duplexer, filter, frequency synthesizer, etc., in order to implement at least one of frequency division duplex (FDD) and time division duplex (TDD). For example, the acquisition unit 11, modification unit 12, estimation unit 13, correction unit 14, and generation unit 15 described above may be implemented by the communication device 1004.
[0094] The input device 1005 is an input device that accepts input from an external source (e.g., a keyboard, mouse, microphone, switch, button, sensor, etc.). The output device 1006 is an output device that outputs to an external source (e.g., a display, speaker, LED lamp, etc.). The input device 1005 and the output device 1006 may be configured as an integrated unit (e.g., a touch panel).
[0095] Furthermore, each device, such as the processor 1001 and memory 1002, is connected by a bus 1007 for communicating information. The bus 1007 may be configured using a single bus, or different buses may be configured for each device.
[0096] Furthermore, the device 1 may be configured to include hardware such as a microprocessor, a digital signal processor (DSP), an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), and an FPGA (Field Programmable Gate Array), and some or all of each functional block may be realized by such hardware. For example, the processor 1001 may be implemented using at least one of these hardware components.
[0097] The notification of information is not limited to the manner / embodiments described herein and may be carried out by other means.
[0098] Each aspect / embodiment described in this disclosure may be applied to at least one of the following systems: LTE (Long Term Evolution), LTE-A (LTE-Advanced), SUPER 3G, IMT-Advanced, 4G (4th generation mobile communication system), 5G (5th generation mobile communication system), FRA (Future Radio Access), NR (new Radio), W-CDMA®, GSM®, CDMA2000, UMB (Ultra Mobile Broadband), IEEE 802.11 (Wi-Fi®), IEEE 802.16 (WiMAX®), IEEE 802.20, UWB (Ultra-WideBand), Bluetooth®, and other appropriate systems, as well as next-generation systems extended based thereon. Furthermore, multiple systems may be applied in combination (for example, a combination of at least one of LTE and LTE-A with 5G).
[0099] The processing procedures, sequences, flowcharts, etc., of each aspect / embodiment described in this disclosure may be reordered, provided they do not contradict each other. For example, the methods described in this disclosure present various step elements using exemplary order and are not limited to the specific order presented.
[0100] Input and output information may be stored in a specific location (e.g., memory) or managed using a management table. Input and output information may be overwritten, updated, or appended to. Output information may be deleted. Input information may be transmitted to other devices.
[0101] The determination may be made by a value represented by one bit (0 or 1), by a boolean value (true or false), or by a numerical comparison (for example, a comparison with a predetermined value).
[0102] Each aspect / embodiment described in this disclosure may be used individually, in combination, or switched between as needed during implementation. Furthermore, notification of specific information (e.g., notification that "X is") is not limited to explicit notification, but may also be implicit (e.g., by not providing such notification).
[0103] Although the present disclosure has been described in detail above, it will be clear to those skilled in the art that the present disclosure is not limited to the embodiments described herein. The present disclosure can be implemented in modified and altered forms without departing from the intent and scope of the present disclosure as defined by the claims. Therefore, the descriptions in the present disclosure are illustrative and not intended to be restrictive in any way.
[0104] Software should be broadly interpreted to mean instructions, instruction sets, code, code segments, program code, programs, subprograms, software modules, applications, software applications, software packages, routines, subroutines, objects, executable files, execution threads, procedures, functions, and so on, whether they are called software, firmware, middleware, microcode, hardware description languages, or by any other name.
[0105] Furthermore, software, instructions, information, etc., may be transmitted and received via a transmission medium. For example, if software is transmitted from a website, server, or other remote source using at least one of wired technology (such as coaxial cable, fiber optic cable, twisted pair, or digital subscriber line (DSL)) and wireless technology (such as infrared or microwave), then at least one of these wired and wireless technologies is included in the definition of a transmission medium.
[0106] The information, signals, etc. described in this disclosure may be represented using any of the various different techniques. For example, the data, instructions, commands, information, signals, bits, symbols, chips, etc. that may be referred to throughout the above description may be represented by voltage, current, electromagnetic waves, magnetic fields or magnetic particles, optical fields or photons, or any combination thereof.
[0107] In addition, terms used in this disclosure and terms necessary for understanding this disclosure may be replaced with terms having the same or similar meaning.
[0108] The terms “system” and “network” as used in this disclosure are interchangeable.
[0109] Furthermore, the information, parameters, etc., described in this disclosure may be expressed using absolute values, relative values from a predetermined value, or corresponding other information.
[0110] The names used for the parameters described above are not restrictive in any way. Furthermore, the formulas and other expressions using these parameters may differ from those expressly disclosed in this disclosure.
[0111] As used in this disclosure, the terms “determining” and “decision” may encompass a wide variety of actions. “Determining” may include, for example, judging, calculating, computing, processing, deriving, investigating, looking up, search, inquiry (e.g., searching in tables, databases or other data structures), and ascertaining. “Determining” may also include receiving (e.g., receiving information), transmitting (e.g., sending information), input, output, and accessing (e.g., accessing data in memory). Furthermore, “determining” may include resolving, selecting, choosing, establishing, and comparing. In other words, "judgment" and "decision" can include considering that some action has been "judged" or "decided." Also, "judgment (decision)" can be reinterpreted as "assuming," "expecting," or "considering."
[0112] The terms “connected,” “coupled,” and any variations thereof mean any direct or indirect connection or coupling between two or more elements, and may include the presence of one or more intermediate elements between two elements that are “connected” or “coupled” with each other. The coupling or connection between elements may be physical, logical, or a combination thereof. For example, “connection” may be reinterpreted as “access.” As used in this disclosure, two elements may be considered to be “connected” or “coupled” with each other using at least one of one or more wires, cables, and printed electrical connections, and, in some non-limiting and non-exclusive examples, electromagnetic energy having wavelengths in the radio frequency domain, microwave domain, and optical (both visible and invisible) domain.
[0113] In this disclosure, the phrase "based on" does not mean "based solely on" unless otherwise specified. In other words, the phrase "based on" means both "based solely on" and "based at least on."
[0114] Any reference to elements using the designations “first,” “second,” etc., as used in this disclosure does not generally limit the quantity or order of those elements. These designations may be used in this disclosure as a convenient way to distinguish between two or more elements. Accordingly, references to the first and second elements do not imply that only two elements may be employed, or that the first element must precede the second element in any way.
[0115] In the configuration of each of the above devices, "means" may be replaced with "part," "circuit," "device," etc.
[0116] Where the terms “include,” “including,” and variations thereof are used in this disclosure, these terms are intended to be inclusive, as is the term “comprising.” Furthermore, the term “or” as used in this disclosure is not intended to mean exclusive OR.
[0117] In this disclosure, if articles are added by translation, such as a, an, and the in English, this disclosure may include the fact that the noun following these articles is plural.
[0118] In this disclosure, the term "A and B are different" may mean "A and B are different from each other." The term may also mean "A and B are each different from C." Terms such as "separate" and "combine" may be interpreted similarly to "different."
[0119] 1...Device, 2...External server, 10...Storage unit, 11...Acquisition unit, 12...Modification unit, 13...Estimation unit, 14...Correction unit, 15...Generation unit, 20...LLM, 1001...Processor, 1002...Memory, 1003...Storage, 1004...Communication device, 1005...Input device, 1006...Output device, 1007...Bus.
Claims
1. An apparatus comprising: an acquisition unit that acquires multiple-choice questions; and an estimation unit that estimates the position bias of the generating AI with respect to the questions based on the probability of appearance of the options of a modified question output by the generating AI in response to a modified question in which all the options of the acquired question have been changed to the same content.
2. The apparatus according to claim 1, wherein the estimation unit estimates based on the difference between the probability of occurrence and the probability of occurrence if all the options in the modified question had the same probability of occurrence.
3. The apparatus according to claim 1, wherein the estimation unit estimates the position bias of the generating AI for one of the options in the question corresponding to one of the options, based on the difference between the probability of appearance of one of the options in the modified question output by the generating AI in response to the modified question and the probability of appearance of all of the options in the modified question having the same probability of appearance.
4. The apparatus according to claim 1, wherein the generating AI is controlled to select one of the options for a multiple-choice question.
5. The apparatus according to claim 1, further comprising a correction unit that corrects the probability of appearance of the options for the question output by the generating AI in response to the question, based on the position bias estimated by the estimation unit.
6. The apparatus according to claim 1, further comprising a modification unit that generates a modified question in which all the options of the question acquired by the acquisition unit have been changed to the same content, and the estimation unit estimates the position bias of the generating AI with respect to the question based on the probability of appearance of the options of the modified question output by the generating AI with respect to the modified question generated by the modification unit.
7. The apparatus according to claim 1, further comprising a generation unit that generates an estimation model for estimating the position bias of the generating AI for a multiple-choice question by learning training data including feature information relating to one or more features of the question and the position bias of the generating AI for the question estimated by the estimation unit.
8. The apparatus according to claim 7, wherein the generation unit learns the feature information contained in the training data as an explanatory variable and the position bias contained in the training data as an objective variable.
9. The apparatus according to claim 7, wherein the estimation unit estimates the position bias of the generating AI for the question by inputting the characteristic information of the question acquired by the acquisition unit into the estimation model generated by the generation unit.
10. A method performed by a computer, comprising: an acquisition step of acquiring a multiple-choice question; and an estimation step of estimating the position bias of the generating AI for the question based on the probability of appearance of the options of the modified question output by the generating AI for a modified question in which all the options of the question acquired in the acquisition step have been changed to the same content.
