Information processing device, evaluation method, and program

Abstract

Provided is an information processing device comprising: an instruction processing unit that inputs, to a language model, a first text which is an authenticity determination target and a second text which instructs an authenticity determination to be performed, and that acquires an authenticity determination result output from the language model; and an evaluation unit that evaluates a hallucination detection performance of the language model on the basis of whether or not the authenticity determination result is correct.

Description

Information processing device, evaluation method, and program

[0001] The present invention relates to a technique for generating sentences using a language model.

[0002] By providing instructions to a language model, such as a large-scale language model, the language model can generate fluent and persuasive responses. However, language models can sometimes produce responses that are not based on facts. This phenomenon is called hallucination.

[0003] Language models are used in many applications, such as automated text creation and information retrieval, and hallucination reduces the reliability of the application, so application developers need to detect hallucination.

[0004] Fabio Petroni, Tim Rockt¨aschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, YuxiangWu, and Alexander Miller. Language models as knowledge bases? In EMNLPIJCNLP, pp. 2463-2473, 2019.

[0005] It is conceivable to use a language model for hallucination detection. To use a language model for hallucination detection, it is necessary to evaluate (estimate) the hallucination detection performance of the language model. However, there is no prior art for evaluating the hallucination detection performance of a language model.

[0006] The present invention has been made in view of the above points, and aims to make it possible to evaluate the hallucination detection performance of a language model.

[0007] According to the disclosed technology, an information processing device is provided that includes: an instruction processing unit that inputs a first text to be judged for truth or falsity and a second text that instructs to perform a truth or falsity judgment into a language model and obtains a truth or falsity judgment result output from the language model; and an evaluation unit that evaluates the hallucination detection performance of the language model based on whether the truth or falsity judgment result is correct.

[0008] The disclosed technology makes it possible to evaluate the hallucination detection performance of a language model.

[0009] 1 is a configuration diagram of an information processing device 100. FIG. 2 is a flowchart for explaining the operation of the information processing device 100. FIG. 3 is a diagram for explaining a third embodiment. FIG. 4 is a diagram for explaining a fourth embodiment. FIG. 5 is a diagram showing an example of a hardware configuration of the information processing device 100.

[0010] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. The embodiment described below is merely an example, and the embodiment to which the present invention is applied is not limited to the following embodiment.

[0011] (Device Configuration) In this embodiment, the information processing device 100 evaluates (estimates) the hallucination detection performance of a language model.

[0012] Fig. 1 shows an example of the configuration of an information processing device 100 according to this embodiment. Fig. 1 also shows a language model 200. Any model may be used as the language model 200 according to this embodiment, as long as it outputs text corresponding to input text. In this embodiment, a pre-trained language model, which is a neural network model, is used as the language model 200.

[0013] 1 , the language model 200 is provided outside the information processing device 100. More specifically, the language model 200 is provided in a computer outside the information processing device 100. The computer may be a physical machine or a virtual machine on the cloud. Note that the language model 200 may also be provided inside the information processing device 100.

[0014] 1, the information processing device 100 includes an instruction processing unit 110 and an evaluation unit 120. The instruction processing unit 110 inputs an instruction statement to a language model 200, and receives from the language model 200 a sentence generated by the language model 200 in response to the instruction statement.

[0015] The evaluation unit 120 evaluates the hallucination detection performance of the language model 200 based on the output from the language model 200 .

[0016] Note that the text (such as a directive or a sentence to be judged for truth or falsehood) input to the language model 200 does not have to be in the form of a sentence, as long as it can convey its intention to the language model 200. For example, the text input to the language model 200 may be a list of words, etc. Furthermore, the text output from the language model 200 does not have to be in the form of a sentence.

[0017] Furthermore, the "truth determination" of text performed by the language model 200 is not limited to a 1 / 0 determination of whether the text is correct or incorrect. For example, the language model 200 may respond with a score indicating the degree of truth or falsehood, and calculation of such a score is also included in the "truth determination."

[0018] Furthermore, the instruction processing unit 110 and the evaluation unit 120 may be provided in separate devices (computers), and a system may be configured with the instruction processing unit 110 and the evaluation unit 120. Even in this case, the device having the instruction processing unit 110 and the evaluation unit 120 is referred to as the information processing device 100.

[0019] Below, examples 1 to 4 will be described as detailed operation examples of the information processing device 100. Examples 3 and 4 correspond to variations of Examples 1 and 2. Note that the overall flow of processing is common to Examples 1 and 2, so both Examples 1 and 2 will be described using the flowchart in FIG.

[0020] First Embodiment The operation of the information processing apparatus 100 in the first embodiment will be described along the steps of the flowchart in FIG.

[0021] <S101 (Step 101)> In S101, the instruction processing unit 110 inputs an instruction statement instructing the generation of a sentence to the language model 200, and obtains a false sentence or a true sentence output by the language model 200 in response to the instruction statement.

[0022] The sentence that the instruction processing unit 110 inputs to the language model 200 may be a sentence that the user inputs to the instruction processing unit 110, or may be a sentence that the instruction processing unit 110 obtains from a database or a predetermined site on the Internet. More specific processing in S101 is as follows.

[0023] First, the instruction processing unit 110 inputs instruction statement 1, an input sentence, and an object phrase into the language model 200, and obtains a true sentence output from the language model 200. In this example, instruction statement 1 is a sentence for causing the language model 200 to generate a true sentence based on the input sentence and the object phrase. It is also possible to cause the language model 200 to generate a true sentence without using the "input sentence and object phrase."

[0024] A true statement is one that contains only true triples, which are real-world facts expressed as (subject, relation, object).

[0025] The instruction processing unit 110 also inputs instruction statement 2, the input sentence, and the object phrase into the language model 200, and obtains a false sentence output from the language model 200. A false sentence is a sentence that includes an incorrect triple. In this example, instruction statement 2 is a sentence that causes the language model 200 to generate a false sentence based on the input sentence and the object phrase. It is also possible to cause the language model 200 to generate a false sentence without using the "input sentence and object phrase."

[0026] In the first embodiment, when the language model 200 is made to generate a false sentence, the instruction processing unit 110 instructs the language model 200 to tell a lie that appears true at first glance, thereby making it possible to generate a false sentence that is not obvious (that is, difficult to detect as false). This instruction corresponds to the part "Generate a false sentence that seems true" in the specific example described later.

[0027] A specific example using GPT-3.5 Turbo (https: / / platform.openai.com / docs / models / gpt-3-5) as the language model 200 is shown below. Note that in Example 1 (and Example 2), a specific example using English is shown, but this is merely an example. The technology according to this embodiment can be applied to any language. (1) Output of true sentences Input to language model 200: Generate a paraphrase of the input. You must use the keyword: Input: NahooToo is the second album by Scottish musician Paul Mounsey. Keyword: Scotland Output from language model 200: Paraphrase: Paul Mounsey released NahooToo as his second album as a musician from Scotland. Explanation of (1): The above "Generate a paraphrase of the input. You must use the keyword:" corresponds to instruction statement 1, "Input: NahooToo is the second album by Scottish musician Paul Mounsey." corresponds to the input sentence, and "Keyword: Scotland" corresponds to the object phrase.(2) Output of a false sentence Input to language model 200: Generate a false sentence that seems true. You must rewrite only the one keyword in the input: Input: Paul Mounsey released NahooToo as his second album as a musician from Scotland. Keyword: Scotland Output from language model 200: False sentence: Paul Mounsey released NahooToo as his second album as a musician from Ireland. Explanation of (2): The above "Generate a false sentence that seems true. You must rewrite only the one keyword in the input:" corresponds to instruction 2, "Input: Paul Mounsey released NahooToo as his second album as a musician from Scotland." corresponds to the input sentence, and "Keyword: Scotland" corresponds to the object phrase.

[0028] <S102> In S102, the instruction processing unit 110 inputs the instruction sentence and the output from the language model 200 acquired in S101 to the language model 200, and acquires the truth / false judgment result output by the language model 200. Note that the instruction sentence may not be used. For example, when a sentence to be judged as true / false is input to the language model 200, the language model 200 may interpret this as a sentence for which a truth / false judgment is to be made, and may then judge the truth / false. The same applies to other embodiments in which the instruction sentence may not be used.

[0029] The instruction processing unit 110 also holds the positive or negative truth determination result (the correct answer as to whether the truth determination result is correct or not), and determines whether the truth determination result output by the language model 200 is correct or not (correct answer / incorrect answer) based on the held information. The correct answer as to whether the truth determination result is correct or not may be input in advance by the user to the instruction processing unit 110, or the instruction processing unit 110 may acquire it by referring to external knowledge or the like.

[0030] That is, in S102, the language model 200 is caused to determine the truth or falsity of the sentence output by the language model 200 in S101. Note that in the first embodiment, the language model that outputs a true sentence / false sentence and the language model that determines the truth or falsity of a true sentence / false sentence are the same language model 200, but this is not limited to this. The language model that outputs a true sentence / false sentence and the language model that determines the truth or falsity of a true sentence / false sentence may be different.

[0031] Furthermore, the instruction processing unit 110 may instruct the language model 200 to use Chain-of-Thought (CoT) in an instruction statement for making the language model 200 determine the truth or falsity of a sentence. This allows the language model 200 to divide the sentence to be determined as the truth or falsity target into predetermined units (each semantic unit) and determine the truth or falsity. Specifically, by instructing the language model 200 to think step by step, a thought process is generated before a determination result is generated, thereby improving the accuracy of the determination. The statement instructing the language model 200 to use CoT corresponds to the part "Let's think step by step to answer this question" in the specific example described below.

[0032] As described above, the instruction processing unit 110 may automatically determine whether the truth determination result by the language model 200 is correct or incorrect based on the correct answer information it holds, or the instruction processing unit 110 may display the truth determination result to the user, allowing the user to determine whether the answer is correct or incorrect.

[0033] The true / false determination result obtained in S102 and information on whether the answer is correct or incorrect are input to the evaluation unit 120. In the first embodiment, S101 and S102 are executed multiple times, and a set of "the true / false determination result and information on whether the answer is correct or incorrect" is accumulated in the evaluation unit 120.

[0034] In Example 1, the "true / false determination results and information on whether the answers are correct or incorrect" for a plurality of false sentences and a plurality of true sentences are acquired as the set. However, this is not limited to this, and the "true / false determination results and information on whether the answers are correct or incorrect" for only a plurality of false sentences may be acquired as the set, or the "true / false determination results and information on whether the answers are correct or incorrect" for only a plurality of true sentences may be acquired as the set.

[0035] Note that the language model 200 forgets past inputs and outputs unless a special setting is made for the language model 200 to remember the previous inputs and outputs. Therefore, in the first embodiment (and the second embodiment), the language model 200 is made to generate a false sentence / true sentence in S101, and in S102, the language model 200 is made to respond as to the truth or falsity of the false sentence / true sentence in a state in which the language model 200 has forgotten it.

[0036] The specific process for causing the language model 200 to execute the truth / false determination in S102 is as follows.

[0037] The instruction processing unit 110 inputs instruction statement 3A (an instruction statement that does not instruct the use of CoT) or instruction statement 4A (an instruction statement that instructs the use of CoT) and the sentence to be judged as true or false into the language model 200, and obtains the true or false judgment result output from the language model 200.

[0038] A specific example when GPT-3.5 Turbo is used as the language model 200 is shown below. (1) When CoT is not used Input to language model 200: Does the input contain misinformation? You must answer with yes or no: Input: Paul Mounsey released NahooToo as his second album as a musician from Scotland. Output from language model 200: Answer: yes Explanation of (1): The above "Does the input contain misinformation? You must answer with yes or no:" corresponds to instruction statement 3A, and "Input: Paul Mounsey released NahooToo as his second album as a musician from Scotland." corresponds to the sentence to be judged as true or false (here, the true sentence obtained in S101). "yes" is the result of the truth judgment. (2) Using CoT Input to language model 200: Does the input contain misinformation? Let's think step by step to answer this question. You must answer with yes or no: Input: Paul Mounsey released NahooToo as his second album as a musician from Ireland. Output from language model 200: Thought: Let's break down the input sentence to evaluate whether it contains misinformation: 1. "Paul Mounsey released NahooToo as his second album": This part of the sentence is accurate."NahooToo" is indeed the second album released by Paul Mounsey. 2. "as a musician from Ireland": This part of the sentence contains misinformation. Paul Mounsey is not from Ireland; he is a Scottish musician and composer. Therefore, the second part of the sentence contains misinformation, where it incorrectly identifies Paul Mounsey as being from Ireland. Answer: yes. Explanation for (2): The above "Does the input contain misinformation? Let's think step by step to answer this question. You must answer with yes or no:" corresponds to instruction 4A, and "Input: Paul Mounsey released NahooToo as his second album as a musician from Ireland." corresponds to the sentence to be judged (here, the false sentence obtained in S101). "yes" is the result of the judgment. The output from language model 200 shows step-by-step thinking.

[0039] <S103> In S103, the evaluation unit 120 calculates a "hallucination detection ability score" that represents the model truth determination ability of the language model 200 that performed the truth determination, based on the set of "truth determination results, correct / incorrect" obtained in S102.

[0040] For example, the evaluation unit 120 calculates at least one of the recall rate, precision rate, F-value, accuracy rate, and error rate as the "hallucination detection ability score" from the set of "truth / falsehood judgment results, correct / incorrect." The recall rate represents the ability to judge falsehood as falsehood, the precision rate represents the ability to judge truth as truth, and the F-value, accuracy rate, and error rate represent the overall ability of these. Specifically, the recall rate, precision rate, F-value, accuracy rate, and error rate can be calculated using the following formulas.

[0041] Recall rate = (number of false sentences whose truthfulness was correctly determined) / (number of false sentences) Precision rate = (number of true sentences whose truthfulness was correctly determined) / (number of true sentences) F-value = 2 x recall rate x precision rate / (recall rate + precision rate) Accuracy rate = (number of sentences whose truthfulness was correctly determined) / ((number of false sentences) + (number of true sentences)) Error rate = 1 - accuracy rate <S104> In S104, the evaluation unit 120 outputs the hallucination detection ability score calculated in S103.

[0042] (Example 2) Next, Example 2 will be described. Example 2 will also be described following the procedure of the flowchart in FIG.

[0043] <S101> S101 is the same as S101 in the first embodiment.

[0044] <S102> In S102, the instruction processing unit 110 inputs the instruction sentence and the output from the language model 200 acquired in S101 to the language model 200, and acquires the truth determination result output by the language model 200 and the confidence level for the truth determination result.

[0045] The instruction processing unit 110 holds the positive or negative truth determination result (the correct answer as to whether the truth determination result is correct or not), and determines whether the truth determination result output by the language model 200 is correct or not (correct / incorrect) based on the held information. The correct answer as to whether the truth determination result is correct or not may be input in advance by the user to the instruction processing unit 110, or the instruction processing unit 110 may acquire it by referring to external knowledge or the like.

[0046] That is, in S102, the language model 200 is made to determine the truth or falsity of the sentence output by the language model 200 in S101, and to output a certainty factor. Note that in the second embodiment, the language model that outputs a true sentence / false sentence and the language model that determines the truth or falsity of a true sentence / false sentence are the same language model 200, but this is not limited to this. The language model that outputs a true sentence / false sentence and the language model that determines the truth or falsity of a true sentence / false sentence may be different.

[0047] In the second embodiment, the instruction processing unit 110 instructs the language model 200 to generate a confidence level for the truth / falseness determination result that it has generated. This instruction corresponds to the part "Then provide the probability that your answer is correct between 0.0 and 1.0" in the specific example described later. In this embodiment, as shown in the score calculation described later, if the determination result is correct, the higher the confidence level, the higher the score, and if the determination result is incorrect, the lower the confidence level, the lower the score.

[0048] Also, in Example 2, as in Example 1, the instruction processing unit 110 may instruct the language model 200 to use a Chain-of-Thought (CoT) in an instruction statement for determining the truth or falsehood of a sentence.

[0049] As described above, the instruction processing unit 110 may automatically determine whether the truth determination result by the language model 200 is correct or incorrect based on the correct answer information it holds, or the instruction processing unit 110 may display the truth determination result to the user, allowing the user to determine whether it is correct or incorrect.

[0050] In Example 2, the instruction processing unit 110 obtains the "true / false determination result, certainty, correct / incorrect" for the true sentence generated in S101, and the "true / false determination result, certainty, correct / incorrect" for the false sentence generated in S101, and inputs these to the evaluation unit 120. The evaluation unit 120 holds the information input from the instruction processing unit 110.

[0051] In addition, the "true / false determination result, certainty, correct / incorrect" for multiple true sentences and the "true / false determination result, certainty, correct / incorrect" for multiple false sentences may be input to the evaluation unit 120, or the "true / false determination result, certainty, correct / incorrect" for only one or multiple true sentences may be input to the evaluation unit 120, or the "true / false determination result, certainty, correct / incorrect" for only one or multiple false sentences may be input to the evaluation unit 120.

[0052] The specific process for causing the language model 200 to execute the truth determination and the certainty factor generation in S102 is as follows.

[0053] The instruction processing unit 110 inputs instruction statement 3B (an instruction statement that does not instruct the use of CoT) or instruction statement 4B (an instruction statement that instructs the use of CoT) and the sentence to be judged as true or false into the language model 200, and obtains the true or false judgment result and confidence level output from the language model 200.

[0054] A specific example when GPT-3.5 Turbo is used as the language model 200 is shown below. (1) When CoT is not used Input to language model 200: Does the input contain misinformation? You must answer with yes or no. Then provide the probability that your answer is correct between 0.0 and 1.0: Input: Paul Mounsey released NahooToo as his second album as a musician from Scotland. Output from language model 200: Answer: yes Probability: 0.9 Explanation of (1): The above "Does the input contain misinformation? You must answer with yes or no. Then provide the probability that your answer is correct between 0.0 and 1.0:" corresponds to instruction 3B, and "Input: Paul Mounsey released NahooToo as his second album as a musician from Scotland." corresponds to the sentence to be judged as true or false (here, the true sentence obtained in S101). "yes" is the truth judgment result, and "0.9" is the confidence level. (2) When using CoT Input to language model 200: Does the input contain misinformation? Let's think step by step to answer this question. You must answer with yes or no. Then provide the probability that your answer is correct between 0.0 and 1.0: Input: Paul Mounsey released NahooToo as his second album as a musician from Ireland.Output from language model 200: Thought: Let's break down the input sentence to evaluate whether it contains misinformation: 1. "Paul Mounsey released NahooToo as his second album": This part of the sentence is accurate. "NahooToo" is indeed the second album released by Paul Mounsey. 2. "as a musician from Ireland": This part of the sentence contains misinformation. Paul Mounsey is not from Ireland; he is a Scottish musician and composer. So, the sentence contains misinformation in the second part where it incorrectly identifies Paul Mounsey as being from Ireland. Answer: yes Probability: 0.9 Explanation for (2): The above "Does the input contain misinformation? Let's think step by step to answer this question. You must answer with yes or no. Then provide the probability that your answer is correct between 0.0 and 1.0:" corresponds to instruction 4B, and "Input: Paul Mounsey released NahooToo as his second album as a musician from "Ireland." corresponds to the sentence to be judged for truth (here, the false sentence obtained in S101). "yes" is the result of the truth judgment, and "0.9" is the confidence level. The output from the language model 200 shows a step-by-step process of thinking.

[0055] <S103> In S103, the evaluation unit 120 calculates a "hallucination detection capability score" that represents the model truth determination capability of the language model 200 that performed the truth determination based on the "truth determination result, confidence level, correct / incorrect" obtained in S102.

[0056] In Example 2, the evaluation unit 120 calculates the average value of the confidence levels of the truth judgment results obtained for one or more sentences as the "hallucination detection ability score." However, if the judgment result is incorrect, the confidence level is multiplied by -1 to calculate the average value. A specific example is described below.

[0057] Assume that the evaluation unit 120 receives one "truth or falsity determination result, certainty, correct / incorrect" for a true sentence and one "truth or falsity determination result, certainty, correct / incorrect" for a false sentence. Also, assume that the truth or falsity determination result for the false sentence is correct with a certainty of 0.9, and that the truth or falsity determination result for the true sentence is incorrect with a certainty of 0.1. In this case, the evaluation unit 120 calculates the hallucination detection ability score using the following formula:

[0058] Hallucination detection ability score = (0.9 + (-0.1)) / 2 = 0.4 In the above example, if the truth judgment result of the false sentence is correct and the confidence level is 0.9, and the truth judgment result of the true sentence is also correct and the confidence level is 0.9, the hallucination detection ability score will be as follows.

[0059] Hallucination detection ability score = (0.9 + 0.9) / 2 = 0.9 Alternatively, the hallucination detection ability score may be calculated solely from the truth determination results of the false sentences. For example, if the truth determination result of false sentence 1 is correct and the certainty is 0.9, and the truth determination result of false sentence 2 is incorrect and the certainty is 0.3, the hallucination detection ability score is as follows:

[0060] Hallucination detection ability score = (0.9 + (-0.3)) / 2 = 0.3 The hallucination detection ability score may also be calculated from only the truth judgment results of the true sentences. For example, if the truth judgment result of true sentence 1 is correct and the certainty is 0.9, and the truth judgment result of true sentence 2 is incorrect and the certainty is 0.5, the hallucination detection ability score is as follows:

[0061] Hallucination detectability score=(0.9+(-0.5)) / 2=0.2 <S104> In S104, the evaluation unit 120 outputs the hallucination detectability score calculated in S103.

[0062] (Other Examples Related to Examples 1 and 2) In Examples 1 and 2, the sentence to be subjected to truth / falsehood judgment (false sentence / true sentence) is generated by the language model 200 (or another language model), but this is not limited to this. For example, a pair of the sentence to be subjected to truth / falsehood judgment and truth / false-answer information may be manually created, and the language model 200 may be instructed to judge the truth / falsehood of the sentence in S102. Alternatively, the sentence to be subjected to truth / falsehood judgment may be acquired from a pre-prepared dataset consisting of a pair of the sentence to be subjected to truth / false-answer information and truth / false-answer information, and the language model 200 may be instructed to judge the truth / falsehood of the sentence in S102.

[0063] Furthermore, in S101 of the first and second embodiments, when instructing the language model 200 to generate a false sentence and a true sentence, the input sentence and the object phrase may not be input to the language model 200. In other words, in S101 of the first and second embodiments, when instructing the language model 200 to generate a false sentence / true sentence, any instruction statement may be used as long as it instructs the language model 200 to generate a false sentence / true sentence.

[0064] (Example 3) Next, Example 3 will be described. Example 3 is a variation that can be applied to both Example 1 and Example 2. In Examples 1 and 2, the instruction processing unit 110 inputs an instruction sentence and a sentence to be evaluated to the language model 200, and the language model 200 returns a truth determination result (and a confidence level in the case of Example 2) to the information processing device 100 (specifically, the instruction processing unit 110), but the present invention is not limited to this form.

[0065] 3, in the third embodiment, the instruction processing unit 110 inputs an instruction sentence and a sentence to be evaluated to the language model 200, and the language model 200 outputs a truth determination result to an output destination other than the information processing device 100. This output destination may be any device that can hold the truth determination result.

[0066] The instruction processing unit 110 acquires the true / false determination result from the output destination device and sends the true / false determination result and correct answer information to the evaluation unit 120. Note that the correct answer information may be held by the instruction processing unit 110 as in Examples 1 and 2, or the correct answer information may be held in the output destination device, and the instruction processing unit 110 may acquire the true / false determination result and correct answer information from the device. The evaluation method in the evaluation unit 120 is as described in Examples 1 and 2.

[0067] (Fourth embodiment) Next, a fourth embodiment will be described. The fourth embodiment is a variation that can be applied to any of the first to third embodiments.

[0068] 4, in the fourth embodiment, a plurality of language models 200-1 to 200-N (N is an integer equal to or greater than 2) are provided. The plurality of language models 200-1 to 200-N may be provided in separate computers, or some of the language models (or all of the plurality of language models 200-1 to 200-N) may be provided in a single computer.

[0069] 4, the instruction processing unit 110 receives, for example, from a user, specification information specifying one or more language models to be evaluated, and an evaluation dataset (a set of pairs of sentences to be evaluated and correct answer information). Here, it is assumed that the specification information specifies multiple language models (specifically, language models 200-1 and 200-2 in FIG. 4).

[0070] Note that the evaluation dataset may not be accepted from a user, etc. For example, the evaluation dataset may be made publicly available on a server on a network, and the instruction processing unit 110 may accept only specification information specifying one or more language models to be evaluated from a user, etc., and acquire the evaluation dataset from the server. Alternatively, the evaluation dataset may be stored in advance in the instruction processing unit 110.

[0071] Next, the instruction processing unit 110 inputs the sentence to be evaluated, which is included in the evaluation dataset, into the language model 200-1 and the language model 200-2 specified by the specification information, and acquires the truth determination result (and the confidence level in the case of the second embodiment) from each language model. When the third embodiment is applied, the truth determination result is acquired from the device to which each language model is output.

[0072] The instruction processing unit 110 sends to the evaluation unit 120 the correct answer information included in the evaluation data set, the truth determination results from the language model 200-1 and the truth determination results from the language model 200-2.

[0073] The evaluation unit 120 evaluates the hallucination detection performance of each of the language models 200-1 and 200-2 using the evaluation methods described in Examples 1 and 2, and outputs the evaluation results for each language model. The destination of the evaluation results may be the provider of the evaluation dataset (e.g., a user).

[0074] The truth of facts may change over time. With the configuration of Example 4, a user can evaluate and compare the performance of any language model service based on their own dataset (especially, what they consider to be the correct answer).

[0075] (Hardware Configuration Example) The information processing device 100 described in this embodiment can be realized, for example, by causing a computer to execute a program. This computer may be a physical computer or a virtual machine on the cloud.

[0076] That is, the information processing device 100 can be realized by using hardware resources such as a CPU and memory built into a computer to execute a program corresponding to the processing performed by the information processing device 100. The program can be recorded on a computer-readable recording medium (such as a portable memory) and can be saved or distributed. The program can also be provided via a network such as the Internet or email.

[0077] Fig. 5 is a diagram showing an example of the hardware configuration of the computer. The computer in Fig. 3 includes a drive device 1000, an auxiliary storage device 1002, a memory device 1003, a CPU 1004, an interface device 1005, a display device 1006, an input device 1007, an output device 1008, and the like, all of which are interconnected via a bus B. The computer may further include a GPU.

[0078] The program that realizes the processing on the computer is provided by a recording medium 1001, such as a CD-ROM or a memory card. When the recording medium 1001 storing the program is set in the drive device 1000, the program is installed from the recording medium 1001 to the auxiliary storage device 1002 via the drive device 1000. However, the program does not necessarily have to be installed from the recording medium 1001, but may be downloaded from another computer via a network. The auxiliary storage device 1002 stores the installed program as well as necessary files, data, etc.

[0079] The memory device 1003 reads and stores a program from the auxiliary storage device 1002 when an instruction to start the program is received. The CPU 1004 realizes functions related to the information processing device 100 in accordance with the program stored in the memory device 1003. The interface device 1005 is used as an interface for connecting to a network, etc. The display device 1006 displays a GUI (Graphical User Interface) or the like according to the program. The input device 1007 is composed of a keyboard, mouse, buttons, a touch panel, etc., and is used to input various operation instructions. The output device 1008 outputs the results of calculations.

[0080] 3, the information held by the instruction processing unit 110 is information held in the memory device 1003 or the auxiliary storage device 1002. Similarly, the information held by the evaluation unit 120 is information held in the memory device 1003 or the auxiliary storage device 1002.

[0081] (Effects of the embodiment) As described above, the technology according to the present embodiment makes it possible to evaluate the hallucination detection performance of a language model. Furthermore, the technology according to the present embodiment has the advantage that manual evaluation of hallucination detection performance is unnecessary, and therefore evaluation can be performed at low cost.

[0082] That is, the information processing device according to the present embodiment provides a specific improvement to the technology for evaluating the hallucination detection performance of a language model, and represents an advancement in the technical field related to the evaluation of language models. Furthermore, the information processing device according to the present embodiment is incorporated into practical applications.

[0083] The following additional notes are provided regarding the above-described embodiments.

[0084] <Additional Notes> (Additional Item 1) An information processing device including a memory and at least one processor connected to the memory, wherein the processor inputs a first text to be subjected to truth or falsity judgment and a second text instructing to perform a truth or falsity judgment into a language model, obtains a truth or falsity judgment result output from the language model, and evaluates the hallucination detection performance of the language model based on whether the truth or falsity judgment result is correct. (Additional Item 2) The processor further obtains a confidence factor for the truth or falsity judgment result from the language model, and evaluates the hallucination detection performance of the language model based on whether the truth or falsity judgment result is correct and on the confidence factor. (Additional Item 3) The information processing device according to Additional Item 1 or 2, wherein the processor instructs the language model to perform a truth or falsity judgment using a chain of thoughts, using the second text. (Additional Item 4) The information processing device according to any one of Additional Items 1 to 3, wherein the first text is a false sentence or a true sentence generated by the processor in the language model. (Supplementary Item 5) The processor instructs the language model to tell a lie that appears true when generating the false sentence in the language model. (Supplementary Item 6) An information processing device comprising: a memory; and at least one processor connected to the memory, wherein the processor: accepts designation information that designates one or more language models to be evaluated and an evaluation dataset, inputs a first text included in the evaluation dataset to the one or more language models designated by the designation information, obtains truth determination results output from each language model, evaluates the hallucination detection performance of the one or more language models based on the truth determination results and truth correct information included in the evaluation dataset, and outputs an evaluation result for each language model.(Supplementary Item 7) An evaluation method executed by an information processing device, comprising the steps of: inputting a first text to be judged as true or false and a second text instructing to perform a true or false judgment into a language model, and obtaining a true or false judgment result output from the language model, and evaluating the hallucination detection performance of the language model based on whether the true or false judgment result is correct. (Supplementary Item 8) A non-transitory storage medium storing a program for causing a computer to function as an information processing device described in any one of Supplementary Items 1 to 6.

[0085] Although the present embodiment has been described above, the present invention is not limited to such a specific embodiment, and various modifications and changes are possible within the scope of the gist of the present invention described in the claims.

[0086] REFERENCE SIGNS LIST 100 Information processing device 110 Instruction processing unit 120 Evaluation unit 200 Language model 1000 Drive device 1001 Recording medium 1002 Auxiliary storage device 1003 Memory device 1004 CPU 1005 Interface device 1006 Display device 1007 Input device 1008 Output device

Claims

1. An information processing device comprising: an instruction processing unit that inputs a first text to be judged as true or false and a second text that instructs to perform a true or false judgment into a language model and obtains a true or false judgment result output from the language model; and an evaluation unit that evaluates the hallucination detection performance of the language model based on whether the true or false judgment result is correct.

2. The information processing device described in claim 1, wherein the instruction processing unit further obtains a degree of confidence for the truth determination result from the language model, and the evaluation unit evaluates the hallucination detection performance of the language model based on whether the truth determination result is correct and the degree of confidence.

3. The information processing device according to claim 1, wherein the instruction processing unit instructs the language model to perform a truth / false determination using a chain of thoughts based on the second text.

4. The information processing device according to claim 1, wherein the first text is a false sentence or a true sentence that the instruction processing unit causes the language model to generate.

5. The information processing device according to claim 4, wherein the instruction processing unit instructs the language model to tell a lie that appears true when causing the language model to generate the false sentence.

6. An information processing device comprising: an instruction processing unit that receives specification information specifying one or more language models to be evaluated and an evaluation dataset, inputs a first text included in the evaluation dataset into the one or more language models specified by the specification information, and obtains truth determination results output from each language model; and an evaluation unit that evaluates the hallucination detection performance of the one or more language models based on the truth determination results and truth correctness information included in the evaluation dataset, and outputs evaluation results for each language model.

7. An evaluation method executed by an information processing device, comprising the steps of: inputting a first text to be judged for truth or falsity and a second text instructing that a truth or falsity judgment be made into a language model, and obtaining a truth or falsity judgment result output from the language model; and evaluating the hallucination detection performance of the language model based on whether the truth or falsity judgment result is correct.

8. A program for causing a computer to function as each unit in the information processing device according to any one of claims 1 to 6.

Images