Information processing device, information processing method, and program

Abstract

Provided is an information processing device in which a first acquisition means acquires a query. A second acquisition means acquires a plurality of texts pertaining to the query. A generation means generates an answer and grounds for the answer from the query and the plurality of texts by using a machine learning model. Here, in case where an answer cannot be found for the query, the generation means generates the reason for being unable to answer from the plurality of texts. The information processing device can assist the decision making of a user by presenting the answer reason.

Description

Information processing device, information processing method, and recording medium

[0001] The present disclosure relates to the art of question answering.

[0002] In question answering, techniques that utilize search results on web pages and language models have been developed. For example, Patent Literature 1 proposes a question answering device that presents answers to questions and the reasons for the answers.

[0003] International Publication WO2022 / 079826

[0004] In the conventional question-answering method, if a question cannot be answered, it is possible to output "answer not possible," but it is difficult to output the reason for the inability to answer. Furthermore, even with Patent Document 1, if a question cannot be answered, it is not always possible to output the reason.

[0005] One object of the present disclosure is to provide an information processing device that, when it is unable to answer a question, is able to output the reason for this.

[0006] In one aspect of the present disclosure, an information processing device comprises: a first acquisition means for acquiring a query; a second acquisition means for acquiring a plurality of texts related to the query; and a generation means for generating an answer and a basis for the answer from the query and the plurality of texts using a machine learning model, wherein the generation means generates a basis for the inability to answer from the plurality of texts when the query cannot be answered.

[0007] In another aspect of the present disclosure, an information processing method includes: acquiring a query; acquiring a plurality of texts related to the query; using a machine learning model to generate an answer and a basis for the answer from the query and the plurality of texts; and, if the query cannot be answered, generating a basis for the inability to answer from the plurality of texts.

[0008] In yet another aspect of the present disclosure, a recording medium records a program that causes a computer to execute the following processes: acquire a query; acquire a plurality of texts related to the query; use a machine learning model to generate an answer and a basis for the answer from the query and the plurality of texts; and, if the query cannot be answered, generate a basis for the inability to answer from the plurality of texts.

[0009] According to the present disclosure, when a question cannot be answered, it is possible to output the reason why.

[0010] 1 is a block diagram showing a schematic configuration of an information processing device according to the present disclosure; FIG. 2 is an example of an output of an answerable question and an answer to that question; FIG. 3 is an example of an output of an unanswerable question and an answer to that question; FIG. 4 is a block diagram showing a hardware configuration of an information processing device; FIG. 5 is a block diagram showing a functional configuration of an information processing device; FIG. 6 shows an example of processing of a text generation model; FIG. 7 shows another example of processing of a text generation model; FIG. 8 is a flowchart of question answering processing; FIG. 9 is an example of a database to be searched; and FIG. 10 is a display example of question answers. FIG. 11 is a block diagram showing a functional configuration of an information processing device according to a second embodiment; and FIG. 12 is a flowchart of processing by the information processing device according to the second embodiment.

[0011] Preferred embodiments of the present disclosure will be described below with reference to the drawings. First Embodiment [Conceptual Description] FIG. 1 shows an overview of an information processing device 100. A user inputs a question (query) to the information processing device 100. Based on the question, the information processing device 100 acquires documents that are likely to contain an answer from a website or the like. The information processing device 100 divides the acquired documents into sentence-based or paragraph-based information (hereinafter also referred to as "passages" or "text"). Then, based on the question and multiple passages, the information processing device 100 generates and outputs an answer to the question and the basis for that answer (hereinafter also referred to as "answer basis"). If the information processing device 100 cannot answer the question, it generates and outputs information indicating that the answer is not possible and the basis for the inability to answer.

[0012] With known techniques, if a question cannot be answered, it is possible to output "unable to answer," but it is not possible to output the reason for the unanswerable answer. With the information processing device 100 of this embodiment, if a question cannot be answered, it is possible to output "unable to answer" along with the reason for the unanswerable answer. This improves the reliability and robustness of the answer, and also allows the user to modify the question to be input based on the reason for the unanswerable answer.

[0013] Next, an example of output of answerable questions, unanswerable questions, and answers will be described.

[0014] (Answerable Questions) FIG. 2 shows an example of an answerable question and an output answer to that question. FIG. 2(A) is an example of a question and a passage. In response to a user's question, "Which country won the most medals at the Winter Olympics?", the information processing device 100 obtains the following three passages P1 to P3. P1: Norway won 39 medals, the most medals ever won at the Winter Olympics. P2: At this year's Winter Olympics, Norway won the most medals ever, making big news. P3: Snowboarding, a newly added sport at the Winter Olympics, became a hot topic.

[0015] FIG. 2B shows an example of an output answer to a question. The information processing device 100 inputs the question and each passage into a text generation model and outputs an answer and evidence. For example, the text generation model extracts a portion of each passage that is likely to be an answer (hereinafter also referred to as an "answer portion") along with a score indicating the likelihood of that portion being an answer (hereinafter also referred to as a "reliability score"), and determines the answer portion with a reliability score equal to or greater than a predetermined threshold as the final answer. The text generation model also outputs the passage that was the source of the final answer as evidence. In FIG. 2B, the text generation model outputs "Norway" as the answer and information indicating passages P1 and P2 as evidence.

[0016] (Unanswerable Questions) FIG. 3 shows an example of an unanswerable question and an output answer to that question. FIG. 3(A) shows an example of a question and a passage. In response to a user's question, "Which country won the most medals at the Winter Olympics?", the information processing device 100 obtains the following three passages P1 to P3. P1: Norway won 39 medals, the most medals ever won at the Winter Olympics. P2: At this year's Winter Olympics, Germany won the most medals ever, making headlines. P3: Snowboarding, a newly added sport at the Winter Olympics, became a hot topic.

[0017] FIG. 3B shows an example of an output answer to a question. The information processing device 100 inputs a question and each passage into a text generation model and outputs an answer and evidence. In FIG. 3B, the text generation model extracts the answer portion "Norway" from passage P1 with a reliability score equal to or greater than a predetermined threshold, and extracts the answer portion "Germany" from passage P2 with a reliability score equal to or greater than a predetermined threshold. In this case, because the answers between passages P1 and P2 conflict, the text generation model cannot determine a single answer. Therefore, the text generation model outputs "Answer Unavailable" as the answer. The text generation model also outputs the text "Conflict between P1 and P2" as evidence, indicating that passages P1 and P2 are in conflict.

[0018] There are multiple patterns in which a question becomes unanswerable. Below are examples of patterns in which a question becomes unanswerable and output examples of the reasons for this. Pattern 1: There is insufficient information in the passage to answer. Output example of the reason: "No information in passages." Pattern 2: There is a contradiction between the passages. Output example of the reason: "Conflict between Passage X and Y." Pattern 3: The question is ambiguous, so the answer cannot be determined. (Example 1) When the question is ambiguous, so the content of the question cannot be uniquely determined. (Example 2) When the question is ambiguous, so multiple answers are possible. Output example of the reason: "Question is too ambiguous." When the question and multiple passages fall into any of the above patterns 1 to 3, the information processing device 100 outputs the answer "unanswerable" and the reason for this.

[0019] 4 is a block diagram showing the hardware configuration of the information processing device 100. As shown in the figure, the information processing device 100 includes an interface (IF) 11, a processor 12, a memory 13, a recording medium 14, and a database (DB) 15.

[0020] The IF 11 communicates with an external device to acquire a question input by a user, and also transmits an answer and the basis for the answer generated by the information processing device 100 to the external device.

[0021] The processor 12 is a computer such as a CPU (Central Processing Unit) and controls the entire information processing device 100 by executing a pre-prepared program. Specifically, the processor 12 may be a CPU, a GPU (Graphics Processing Unit), a DSP (Digital Signal Processor), an MPU (Micro Processing Unit), an FPU (Floating Point number Processing Unit), a PPU (Physics Processing Unit), a TPU (Tensor Processing Unit), a quantum processor, a microcontroller, or a combination thereof. The processor 12 executes a question-answering process, which will be described later.

[0022] The memory 13 is composed of a ROM (Read Only Memory), a RAM (Random Access Memory), etc. The memory 13 stores various programs executed by the processor 12. The memory 13 is also used as a working memory while the processor 12 is executing various processes.

[0023] The recording medium 14 is a non-volatile, non-transitory recording medium such as a disk-shaped recording medium or a semiconductor memory, and is configured to be detachable from the information processing device 100. The recording medium 14 records various programs to be executed by the processor 12. When the information processing device 100 executes various processes, the programs recorded on the recording medium 14 are loaded into the memory 13 and executed by the processor 12.

[0024] The DB 15 stores a text generation model and the like used in the question answering process described later.

[0025] In addition to the above, the information processing device 100 may also include a display device such as a liquid crystal display or a projector, and an input device such as a keyboard or a mouse. These display devices and input devices are used by, for example, an administrator of the information processing device 100 to perform necessary management.

[0026] 5 is a block diagram showing the functional configuration of the information processing device 100 according to the first embodiment. The information processing device 100 functionally includes an acquisition unit 111, a search unit 112, a text generation unit 113, and an output unit 114.

[0027] A user inputs a question to the information processing device 100. The acquisition unit 111 acquires the question input by the user and outputs it to the search unit 112 and the text generation unit 113.

[0028] The search unit 112 acquires multiple passages from information published on websites based on the user's question. For example, the search unit 112 creates a query for document search based on the user's question. The search unit 112 performs a web search based on the created query and acquires a group of texts that are likely to contain the answer. The search unit 112 divides the group of texts into multiple passages and outputs the multiple passages to the text generation unit 113.

[0029] The text generation unit 113 acquires a question from the search unit 112 and multiple passages from the search unit 112. The text generation unit 113 generates an answer and a basis for the answer from the question and multiple passages using a text generation model described below. The text generation model is an example of a machine learning model. If the question cannot be answered, the text generation model generates "unable to answer" as the answer and generates the basis for the unanswerable answer as the basis for the answer. The text generation unit 113 outputs the answer and the basis for the answer to the output unit 114.

[0030] The output unit 114 outputs the answer and the basis for the answer input from the text generation unit 113 .

[0031] In the above configuration, the acquisition unit 111 is an example of a first acquisition means, the search unit 112 is an example of a second acquisition means, and the text generation unit 113 and the output unit 114 are examples of a generation means.

[0032] [Text Generation Method] Next, a text generation method using a text generation model will be described. In this embodiment, a text generation model based on Fusion-In-Decoder (hereinafter also referred to as "FiD") is used.

[0033] Details of FiD are described in, for example, the following literature, which is incorporated herein by reference: Gautier Izacard, Edouard Grave. "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering," arXiv 2020.

[0034] (Method 1) In method 1, the text generation model outputs both an answer and the basis for the answer. Fig. 6 shows an example of processing by the text generation model. The text generation model in Fig. 6 includes encoders 21a, 21b, and 21c, a combining unit 22, and a decoder 23. Note that when individual encoders are to be distinguished, a subscript is added to the encoder 21, and when they are not to be distinguished, they are simply referred to as "encoder 21."

[0035] The encoder 21 receives a question-passage pair as input and calculates an embedded representation. The combiner 22 concatenates the individual embedded representations calculated by each encoder and outputs the concatenated representation to the decoder 23. The decoder 23 receives the concatenated embedded representation as input and generates and outputs an answer and the basis for the answer. In this case, the decoder 23 may first output the answer and then output the basis for the answer. Alternatively, the decoder 23 may first output the basis for the answer and then output the answer.

[0036] A text generation model is created, for example, by fine-tuning a large-scale language model (hereinafter also referred to as a "language model") that has been pre-trained with a large amount of text data. Specifically, the language model is trained using a dataset of combinations of a question and multiple passages, an answer, and reasons for the answer, and parameters are adjusted so that the language model can output an answer and reasons for the answer from the question and multiple passages.

[0037] Furthermore, the answer and the basis for the answer may be generated by using few-shot learning or the like. Specifically, the answer and the basis for the answer are generated by inputting an input sentence called a prompt into the language model. The prompt includes, for example, the following example sentences and instruction sentences. Example sentence 1: When question 1 and P1, P2, and P3 are given, "Evidence: P1" and "Answer: Answer possible" are output. Example sentence 2: When question 2 and P4, P5, and P6 are given, "Evidence: Conflict between P4 and P5" and "Answer: Answer impossible" are output. Instruction sentence: When question 3 and P7, P8, and P9 are given?

[0038] A language model is a model that learns the relationships between words in a sentence and is a machine learning model that generates related strings from a target string. By using a language model that has learned sentences and phrases from various contexts, it is possible to generate related strings with appropriate content that are related to the target string. For example, a case where a language model is used in question answering will be described. The language model receives an input question such as "What kind of country is Japan?" as the target string. The language model generates a string such as "Japan is an island country in the Northern Hemisphere..." as an answer to the question.

[0039] The learning method of the language model is not particularly limited, but as an example, the language model may be trained to output at least one sentence including an input string. As a specific example, the language model is a Generative Pretrained Transformer (GPT) that outputs a sentence including an input string by predicting a string that is likely to follow the input string. Other examples of language models include T5 (Text-to-Text Transfer Transformer), BERT (Bidirectional Encoder Representations from Transformers), RoBERTa (Robustly optimized BERT approach), and ELECTRA (Efficiently Learning an Encoder that Classifies Token Replacements Accurately).

[0040] Furthermore, the character strings generated by the language model are not limited to natural languages. For example, the language model may output an artificial language (such as a mathematical formula or program source code) in response to a character string input in a natural language. For example, the language model may accept an input question such as, "How do I retrieve data containing a specific character string from a database?" as a target character string. The language model may output program source code for performing database processing. Alternatively, the language model may output a natural language corresponding to a character string input in an artificial language. Furthermore, the content generated by the language model is not limited to character strings. For example, the language model may generate image data, video data, audio data, or other data formats corresponding to the input character string.

[0041] (Method 2) In Method 2, separate units output the answer and the basis for the answer. FIG. 7 shows another example of processing by a text generation model. In FIG. 7, a text generation model 31 and a basis determiner 32 are used in combination. The text generation model 31 generates the answer, and the basis determiner 32 generates the basis for the answer. The text generation model 31 is realized by the above-mentioned FiD.

[0042] In FIG. 7(A), the basis discriminator 32 uses a trained model to output answer basis information from a question and multiple passages. This trained model is assumed to have been trained in advance using a data set that combines a question, multiple passages, and correct answer labels for the answer basis information, for example. Alternatively, as shown in FIG. 7(B), the basis discriminator 32 may receive a question, multiple passages, and an answer generated by the text generation model 31 as input and output answer basis information. In this case, the basis discriminator 32 uses a trained model that has been trained in advance using a data set that combines a question, multiple passages, an answer, and correct answer labels for the answer basis information, for example. In FIG. 7(B), the basis discriminator 32 outputs answer basis information taking into account the answer generated by the text generation model 31, thereby ensuring consistency between the answer generated by the text generation model 31 and the answer basis information generated by the basis discriminator 32.

[0043] (Method 3) In Method 3, the answer and the basis for the answer are also output by separate units. FIG. 8 shows another example of processing by a text generation model. In FIG. 8, a text generation model 41 and a basis discriminator 42 are used in combination. The text generation model 41 generates the answer and a confidence score, and the basis discriminator 42 generates the basis for the answer. The text generation model 41 is realized by the above-mentioned FiD.

[0044] Specifically, the text generation model 41 receives a question and multiple passages as input and generates an answer and a confidence score. Specifically, as shown in FIG. 8A, the text generation model 41 first receives passages one by one and generates an answer and a confidence score for each passage. Furthermore, as shown in FIG. 8B, the text generation model 41 receives passages two by two and generates an answer and a confidence score for each passage. Note that the text generation model 41 may receive a combination of three or more passages as input, in addition to the examples shown in FIGS. 8A and 8B, and generate an answer and a confidence score.

[0045] The basis discriminator 42 receives the answer and the confidence score generated by the text generation model 41 as input and generates an answer basis. The basis discriminator 42 can generate an answer basis based on the difference between the answer and the confidence score due to the difference in the input passage. FIG. 8C shows an example of input / output of the basis discriminator 42. The basis discriminator 42 receives answer data 43 from the text generation model 41. The answer data 43 includes the passage input to the text generation model 41, and the answer and confidence score generated by the text generation model 41. For example, [ALL, Unable to answer, 0.5] indicates that the question and all passages are input to the text generation model 41, and the text generation model 41 generates the answer "Unable to answer" and the confidence score "0.5." Also, [P1, Answerable, 0.8] indicates that the question and passage P1 were input to the text generation model 41, and the text generation model 41 generated "Answerable" as the answer and "0.8" as the confidence score.

[0046] The basis discriminator 42 generates answer grounds from the answer data 43 using a trained model. This trained model is assumed to have been trained in advance using a data set that includes a combination of a question, multiple passages, answers, confidence scores, and correct answer labels for answer grounds. By using the answers and confidence scores obtained by inputting each passage, the basis discriminator 42 can consider the degree to which each passage contributes to the answer and how important it is. Furthermore, by using the answers and confidence scores obtained by inputting two passages, the basis discriminator 42 can consider whether there is a contradiction between the two passages. For example, if the answer obtained by inputting two passages is "unanswerable" and the confidence score at that time is lower than a predetermined value, the basis discriminator 42 can output an answer ground indicating that there is a contradiction between the two passages.

[0047] [Question Answering Process] Next, the question answering process will be described. Fig. 9 is a flowchart of the question answering process by the information processing device 100. This process is realized by the processor 12 shown in Fig. 4 executing a program prepared in advance and operating as each element shown in Fig. 5.

[0048] First, the acquisition unit 111 acquires a question input by a user (step S11) and outputs it to the search unit 112 and the text generation unit 113. Next, the search unit 112 acquires multiple passages based on the question from information published on websites, etc. (step S12). The search unit 112 outputs the multiple passages to the text generation unit 113.

[0049] Next, the text generation unit 113 generates an answer and a basis for the answer based on the question input from the search unit 112 and the multiple passages input from the search unit 112 (step S13). The text generation unit 113 generates an answer and a basis for the answer from the question and the multiple passages using a text generation model. Note that if the question cannot be answered, the text generation model generates "unable to answer" as the answer and the basis for the unanswerable answer as the basis for the answer. The text generation unit 113 outputs the answer and the basis for the answer to the output unit 114. Next, the output unit 114 outputs the answer and the basis for the answer input from the text generation unit 113 (step S14). Then, the processing ends.

[0050] [Modifications] Next, a description will be given of modifications of the first embodiment. The following modifications can be applied to the first embodiment in appropriate combinations.

[0051] (Variation 1) In the first embodiment described above, a text generation model based on FiD is used. However, instead, a text generation model based on other language models such as GPT2 (Generative Pre-trained Transformer) or GPT3 may be used.

[0052] (Variation 2) In the first embodiment described above, the information processing device 100 acquires passages based on questions input by the user, but the method of acquiring passages is not limited to the above. For example, the user may specify a search range for passages. In this case, the information processing device 100 acquires passages based on the question and the search range specified by the user. Furthermore, the information processing device 100 may output answers and reasons for the answers using passages prepared in advance by the user.

[0053] (Modification 3) The information processing device of the above embodiment may be applied to the medical or healthcare field.

[0054] Research and development is constantly taking place in the fields of medicine and healthcare, and existing information is frequently updated as new research and clinical data emerge. As a result, there is a mixture of old and new information on the web, which can lead to conflicting conclusions when answering questions.

[0055] In contrast, the information processing device of the above embodiment displays a message that an answer is not possible and the reason for this when there is a contradiction in the information source, etc. Therefore, the user can make a final decision using the reason presented by the information processing device.

[0056] For example, a user such as a doctor inputs a medical question into an information processing device. Based on the question, the information processing device retrieves multiple passages from a database operated or provided by a highly reliable institution or organization. Figure 10 shows an example of a database to be searched. Based on the question and the multiple passages, the information processing device outputs an answer and the basis for the answer.

[0057] FIG. 11 shows an example of a question and answer display by an information processing device. A display screen 50 includes a question area 51 and an answer area 52. The question area 51 is an area where a user inputs a question. The answer area 52 is an area where an answer to the question and the basis for the answer are displayed. In FIG. 11, the answer area 52 displays "Unable to answer" as the answer. Furthermore, as the evidence, text is displayed indicating that a contradiction exists between the answer derived from document 1 and the answer derived from document 2.

[0058] As described above, by providing answers and the basis for the answers to questions from the user, the user can make a final decision based on the basis provided by the information processing device and their own expertise and experience. Furthermore, the user can evaluate the output results of the information processing device.

[0059] 12 is a block diagram showing the functional configuration of an information processing apparatus according to Embodiment 2. The information processing apparatus 200 includes a first acquisition unit 201, a second acquisition unit 202, and a generation unit 203.

[0060] 13 is a flowchart of processing by the information processing apparatus of the second embodiment. The first acquisition unit 201 acquires a query (step S201). The second acquisition unit 202 acquires multiple texts related to the query (step S202). The generation unit 203 generates an answer and a basis for the answer from the query and the multiple texts using a machine learning model (step S203). Here, if the query cannot be answered, the generation unit 203 generates a basis for the inability to answer from the multiple texts.

[0061] According to the information processing device 200 of the second embodiment, when a question cannot be answered, it is possible to output the reason for this. Therefore, the information processing device 200 allows the user to make a decision based on the reason.

[0062] Some or all of the above-described embodiments can be described as, but are not limited to, the following supplementary notes.

[0063] (Supplementary Note 1) An information processing device comprising: a first acquisition means for acquiring a query; a second acquisition means for acquiring a plurality of texts related to the query; and a generation means for generating an answer and a basis for the answer from the query and the plurality of texts using a machine learning model, wherein the generation means generates a basis for the inability to answer from the plurality of texts when the query cannot be answered.

[0064] (Supplementary Note 2) The information processing device according to Supplementary Note 1, wherein the generation means, when there is no text including an answer to the query, generates information indicating that the answer is not possible as the answer, and generates information indicating that there is insufficient information as the basis for the answer.

[0065] (Supplementary Note 3) The information processing device according to Supplementary Note 1, wherein the generating means outputs information indicating that an answer cannot be determined due to a contradiction between texts, and generates information indicating the text causing the contradiction as the basis for the answer.

[0066] (Supplementary Note 4) The information processing device according to Supplementary Note 1, wherein the generating means outputs information indicating that the answer is impossible as the answer when the question is ambiguous and an answer cannot be determined from the text, and generates information indicating that the question is ambiguous as the basis for the answer.

[0067] (Supplementary Note 5) The information processing device according to Supplementary Note 1, wherein the machine learning model is trained using a dataset including a query, a plurality of texts, an answer, and a basis for the answer, and the answer and the basis for the answer include an inability to answer and a basis for the inability to answer.

[0068] (Supplementary Note 6) The information processing device according to Supplementary Note 1, wherein the machine learning model includes a text generation model and a basis discriminator, the text generation model generates an answer from the query and the plurality of texts, and the basis discriminator generates answer basis based on the answer, the query, and the plurality of texts.

[0069] (Supplementary Note 7) The information processing device according to Supplementary Note 1, wherein the machine learning model includes a text generation model and a basis discriminator, the generation means inputs one piece of text into the text generation model and generates an answer and a confidence score, respectively, and inputs two pieces of text into the text generation model and generates an answer and a confidence score, respectively, and the basis discriminator generates answer basis based on the multiple answers and confidence scores generated by the generation means.

[0070] (Supplementary Note 8) An information processing method comprising: acquiring a query; acquiring a plurality of texts related to the query; using a machine learning model to generate an answer and a basis for the answer from the query and the plurality of texts; and, if the query cannot be answered, generating a basis for the inability to answer from the plurality of texts.

[0071] (Supplementary Note 9) A recording medium having recorded thereon a program that causes a computer to execute the following processes: acquire a query; acquire multiple texts related to the query; use a machine learning model to generate an answer and a basis for the answer from the query and the multiple texts; and, if the query cannot be answered, generate a basis for the inability to answer from the multiple texts.

[0072] Although the present disclosure has been described above with reference to the embodiments and examples, the present disclosure is not limited to the above-described embodiments and examples. Various modifications that can be understood by a person skilled in the art can be made to the configuration and details of the present disclosure within the scope of the present disclosure.

[0073] 100 Information processing device 111 Acquisition unit 112 Search unit 113 Text generation unit 114 Output unit

Claims

1. A first means of obtaining the query, A second acquisition means for obtaining multiple texts related to the aforementioned query, A generation means that generates an answer and the basis for the answer from the query and the plurality of texts using a machine learning model, Equipped with, The generation means is an information processing device that generates the basis for the inability to answer the query from the plurality of texts when it is not possible to answer the query.

2. The information processing apparatus according to claim 1, wherein the generation means generates information indicating that it is impossible to answer the query if there is no text containing an answer to the query, and generates information indicating that there is insufficient information to support the answer.

3. The information processing apparatus according to claim 1, wherein if the answer cannot be determined due to inconsistencies between the texts, the generating means outputs information indicating that an answer is not possible, and generates information indicating the texts that are inconsistent as the basis for the answer.

4. The information processing apparatus according to claim 1, wherein if the generating means cannot determine an answer from the text because the question is ambiguous, it outputs information indicating that it is impossible to answer as the answer, and generates information indicating that the question is ambiguous as the basis for the answer.

5. The aforementioned machine learning model is trained using a dataset that includes queries and multiple texts, as well as answers and justifications for those answers. The information processing apparatus according to claim 1, wherein the aforementioned answer and the basis for the answer include an inability to answer and the basis for the inability to answer.

6. The aforementioned machine learning model comprises a text generation model and a reasoning discriminator. The text generation model generates a response from the query and the multiple texts, The information processing device according to claim 1, wherein the basis discriminator generates the basis for the answer based on the answer, the query and the plurality of texts.

7. The aforementioned machine learning model comprises a text generation model and a reasoning discriminator. The generation means inputs text one by one into the text generation model and generates a response and a confidence score, and inputs two texts at a time into the text generation model and generates a response and a confidence score, The information processing apparatus according to claim 1, wherein the basis discriminator generates the basis for an answer based on a plurality of answers and confidence scores generated by the generation means.

8. A computer-based information processing method, Get the query, Obtain multiple texts related to the aforementioned query, Using a machine learning model, the system generates answers and justifications for those answers from the query and the multiple texts. An information processing method that generates the basis for the inability to answer the query from the multiple texts if the query cannot be answered.

9. Get the query, Obtain multiple texts related to the aforementioned query, Using a machine learning model, the system generates answers and justifications for those answers from the query and the multiple texts. A program that, if it is impossible to answer the aforementioned query, causes the computer to perform a process to generate the basis for the inability to answer from the multiple texts.