Inference method, device and equipment, readable storage medium and program product

Abstract

The invention discloses a reasoning method, device and equipment, a readable storage medium and a program product, and is applied to the technical field of intelligent reasoning, and the method comprises the steps: obtaining an initial reasoning result corresponding to each language based on a target prompt template and a to-be-reasoned problem by using a first large language model; the target prompt template is a template comprising reasoning problem logic; based on the initial reasoning results corresponding to the languages and the multi-choice-question prompt template, obtaining multi-choice questions; the multi-choice question prompt template is a template for constructing multi-choice question logic based on the initial reasoning result; and performing reasoning analysis by using the second large language model based on the multiple choices to obtain a final reasoning result. According to the method, the multi-choice questions are constructed through the reasoning results corresponding to the languages, so that context analysis can be carried out based on the multi-language reasoning results to obtain the final reasoning result, and due to the fact that the final reasoning result can integrate multi-language reasoning, the limitation of current language model reasoning can be reduced, and the reasoning accuracy can be improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent reasoning technology, and particularly to a reasoning method, device, equipment, readable storage medium and program product. Background Art

[0002] Currently, the research on multilingual large language models mainly focuses on how to improve the model's understanding and reasoning abilities between different languages. When performing reasoning currently, a large language model is mainly used to reason about the problems to be reasoned in various languages to obtain reasoning results. However, this method usually faces challenges brought by grammar, context, and cultural differences between languages, resulting in the technical problem of low reasoning accuracy when the model processes complex reasoning tasks.

[0003] It can be seen that how to improve the accuracy of reasoning is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a reasoning method, device, equipment and computer-readable storage medium, which solves the technical problem of low reasoning accuracy in the prior art.

[0005] To solve the above technical problem, the present invention provides a reasoning method, including:

[0006] Using a first large language model based on a target prompt template and a problem to be reasoned to obtain an initial reasoning result corresponding to each language; wherein, the target prompt template is a template including the logic of the reasoning problem;

[0007] Based on the initial reasoning result corresponding to each language and a multiple-choice question prompt template, a multiple-choice question is obtained; wherein, the multiple-choice question prompt template is a template including the logic of constructing a multiple-choice question based on the initial reasoning result;

[0008] Performing reasoning analysis on the multiple-choice question using a second large language model to obtain a final reasoning result.

[0009] On the one hand, using a first large language model based on a target prompt template and a problem to be reasoned to obtain an initial reasoning result corresponding to each language, including:

[0010] Obtaining the problem to be reasoned and the target language corresponding to the problem to be reasoned;

[0011] Based on the problem to be reasoned and the target language, using a target language prompt template to obtain a first input information;

[0012] Based on the problem to be reasoned and the remaining languages, using the language prompt templates corresponding to the remaining languages to obtain second input information; the target prompt template includes the target language prompt template and the language prompt templates;

[0013] Based on the first input information and the second input information, use the first large language model to obtain the initial inference results corresponding to each language.

[0014] On the one hand, based on the initial inference results corresponding to each language and the multiple-choice question prompt template, obtain multiple-choice questions, including:

[0015] Based on the initial inference results corresponding to each language, obtain the inference process and the initial inference results corresponding to each language;

[0016] Use the multiple-choice question objective, and use the inference process and the initial inference results corresponding to each language as options to construct the multiple-choice questions.

[0017] On the other hand, after performing inference analysis on the multiple-choice questions using the second large language model to obtain the final inference results, it further includes:

[0018] Determine the target language corresponding to the question to be inferred;

[0019] Determine whether the final inference results include the inference results corresponding to the target language;

[0020] When the final inference results include the inference results corresponding to the target language, send the inference results corresponding to the target language so that the language of the feedback inference results is consistent with the target language.

[0021] On the other hand, after determining whether the final inference results include the inference results corresponding to the target language, it further includes:

[0022] When the final inference results do not include the inference results corresponding to the target language, use the translation model to translate the final inference results based on the target language to obtain the translated inference results;

[0023] Send the translated inference results.

[0024] On the other hand, when the final inference results do not include the inference results corresponding to the target language, after using the translation model to translate the final inference results based on the target language to obtain the translated inference results, it further includes:

[0025] Perform semantic detection on the translated inference results to determine whether the semantics is accurate;

[0026] When it is accurate, send the translated inference results;

[0027] When it is inaccurate, perform inference analysis based on the final inference result and the translated inference result to obtain the corrected translated inference result.

[0028] On the one hand, after performing inference analysis on the multiple-choice question using the second large language model to obtain the final inference result, it further includes:

[0029] Determine the confidence threshold corresponding to the final inference result;

[0030] On the one hand, the training process of the first large language model includes:

[0031] Train the large language model using a multilingual dataset to obtain an initial large language model; wherein, the data in the multilingual dataset includes news, encyclopedias, papers, and social media texts;

[0032] Fine-tune the initial large language model using a task question set to obtain the first large language model.

[0033] On the one hand, the training process of the second large language model includes:

[0034] Obtain a multilingual multiple-choice question dataset; wherein, the multilingual multiple-choice question dataset includes distraction options, and the distraction options include multimodal distraction data, and the multimodal distraction data includes mixed unit distraction and mixed semantic distraction;

[0035] Train the multilingual pre-trained model using the multilingual multiple-choice question dataset to obtain the second large language model.

[0036] On the one hand, the multiple-choice question prompt template includes question information about the options and each option, and each option is composed of the initial inference result.

[0037] On the one hand, the question to be inferred includes at least one of intelligent education questions and medical diagnosis questions.

[0038] An embodiment of the present invention also provides an inference device, including:

[0039] An initial inference module, configured to use the first large language model based on a target prompt template and a question to be inferred to obtain an initial inference result corresponding to each language; wherein, the target prompt template is a template including the logic of the inference question;

[0040] A multiple-choice question determination module, configured to obtain a multiple-choice question based on the initial inference result corresponding to each language and a multiple-choice question prompt template; wherein, the multiple-choice question prompt template is a template including the logic of constructing a multiple-choice question based on the initial inference result;

[0041] A final reasoning module for performing reasoning analysis on the multiple-choice question using a second large language model to obtain a final reasoning result.

[0042] An embodiment of the present invention also provides a reasoning device, including:

[0043] A memory for storing a computer program;

[0044] A processor for executing the computer program to implement the steps of the reasoning method as described above.

[0045] An embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the reasoning method as described above are implemented.

[0046] An embodiment of the present invention also provides a computer program product, including a computer program / instructions, and when the computer program / instructions are executed by a processor, the steps of the above-mentioned reasoning method are implemented.

[0047] To solve the above technical problems, an embodiment of the present invention provides a reasoning method, including: obtaining an initial reasoning result corresponding to each language by using a first large language model based on a target prompt template and a question to be reasoned; wherein, the target prompt template is a template including the logic of the reasoning question; obtaining a multiple-choice question based on the initial reasoning result corresponding to each language and a multiple-choice question prompt template; wherein, the multiple-choice question prompt template is a template including the logic of constructing a multiple-choice question based on the initial reasoning result; performing reasoning analysis on the multiple-choice question by using a second large language model to obtain a final reasoning result.

[0048] It can be seen from the above technical solutions that the beneficial effect of the present invention is that: compared with the current reasoning process that depends on the prompt words of a certain specific language, resulting in low reasoning accuracy, the present invention constructs a multiple-choice question based on the reasoning results corresponding to each language, so that context analysis can be performed based on the reasoning results of multiple languages to obtain a final reasoning result. Since the final reasoning result can synthesize the reasoning of multiple languages, the limitations of single-language model reasoning can be reduced, so the reasoning accuracy can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] To more clearly illustrate the embodiments of the present invention, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0050] Figure 1 A flowchart of a reasoning method provided by an embodiment of the present invention;

[0051] Figure 2 A flowchart example of a first large language model training method provided by an embodiment of the present invention;

[0052] Figure 3 A flowchart example of an inference method provided by an embodiment of the present invention;

[0053] Figure 4 A structural schematic diagram of an inference device provided by an embodiment of the present invention;

[0054] Figure 5 A structural schematic diagram of an inference device provided by an embodiment of the present invention. Detailed implementation manners

[0055] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0056] The terms "including" and "having" in the specification of the present invention and the accompanying drawings above, and any variations related to "including" and "having", are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may include steps or units not listed.

[0057] In order to enable those skilled in the art of the present technology to better understand the solution of the present invention, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0058] Next, an inference method provided by an embodiment of the present invention will be introduced in detail. Figure 1 A flowchart of an inference method provided by an embodiment of the present invention, the method may include:

[0059] S101, using a first large language model based on a target prompt template and a problem to be inferred, to obtain an initial inference result corresponding to each language; wherein, the target prompt template is a template including the logic of the inference problem.

[0060] The execution entity of this embodiment is an electronic device. This embodiment does not limit the specific electronic device. For example, the electronic device in this embodiment can be a computer, a mobile phone, etc. The problems to be inferred in this embodiment can include at least one of intelligent education problems and medical diagnosis problems, that is, this solution can be applied to the fields of intelligent education technology and medical diagnosis technology. In addition, it can also be applied to problems in other fields, such as the field of mathematics, the field of aerospace, etc. The target prompt template of this embodiment can include: the target to be inferred and the problem to be inferred. For example, it can be "According to the following question, give the correct answer in Chinese: If a cat's speed is 10 kilometers per hour, how much time does it take for it to run 15 kilometers?" or "Based on the following question, provide the correct answer in English: If a cat's speed is 10 kilometers per hour, how much time does it take for it to run 15 kilometers?". This embodiment does not limit the specific solution for obtaining the inference results corresponding to each language. For example, this embodiment can use the same target prompt template for each language; or it can use different target prompt templates for each language, where the form of each prompt template can be the inference target and the problem to be inferred in the corresponding language form, or it can be the inference target in the corresponding language form (the inference target is consistent with the current language) and the problem to be inferred in the corresponding language form (referring to translating the problem to be inferred into the same language as the current language). This embodiment does not limit the specific first large language model. For example, the first large language model in this embodiment can be mBERT (Multilingual BERT), XLM-R (Cross-lingual Model - RoBERTa), etc. The multilingual dataset used by the first large language model in this embodiment includes but is not limited to large-scale corpora such as news, encyclopedias, and social media texts in various languages (the multilingual dataset is a corpus composed of texts in multiple languages, covering natural language data in different languages (such as English, Chinese, Spanish, etc.). Through the training of multilingual corpora, the model can master the basic grammar structures, common vocabulary, and semantic relationships between languages, providing support for cross-lingual inference and problem understanding. As an alternative solution, existing multilingual open-source pre-trained models can also be directly used. After completing multilingual pre-training, the present invention further fine-tunes and trains the large language model by using a multilingual downstream task question set (various inference questions). Through this fine-tuning method, the present invention enables the model to not only be general in terms of language but also show high accuracy in specific downstream task fields.

[0061] It should be further noted that, based on any of the above embodiments, in order to avoid language bias, obtaining the initial inference results corresponding to each language by using the first large language model based on the target prompt template and the problem to be inferred may include:

[0062] S1011, obtaining a problem to be reasoned and a target language corresponding to the problem to be reasoned;

[0063] S1012, obtaining first input information based on the problem to be inferred and the target language using a target language prompt template;

[0064] S1013, based on the problem to be inferred and the remaining language, using the language prompt template corresponding to the remaining language, obtaining second input information; the target prompt template includes a target language prompt template and a language prompt template;

[0065] S1014, obtaining initial inference results corresponding to each language using the first large language model based on the first input information and the second input information.

[0066] This embodiment is different from the traditional multi-language reasoning method. The present invention does not forcibly convert the question to be reasoned input by the user into a form consistent with each language, but uses the original language of the question to be reasoned, thereby avoiding possible translation errors or semantic deviations. For example, first identify the language of the question input by the user, and pass the question to the string variable question. For example, the question input by the user is: "If a cat's speed is 10 kilometers per hour, how long does it take to run 15 kilometers?", then the system analyzes the input content through the language recognition module, identifies that the language of the question is Chinese, and question = "If a cat's speed is 10 kilometers per hour, how long does it take to run 15 kilometers?" By looping through multiple target languages, the questions are passed to the corresponding target language prompt word templates respectively, and a complete model input is obtained, and the reasoning results of the target language version are generated and saved in the candidate result answers list. For example, the system may select the following target languages ​​(depending on which languages ​​have been used in the pre-training and fine-tuning stages of the model): Chinese, English, and call the specific prompt word template for each language version. For example, the Chinese prompt word template is: "Based on the following question, provide the correct answer in Chinese: {question}"; the English prompt word template is: "Based on the following question, provide the correct answer in English: {question}", and the corresponding complete inputs are "Based on the following question, provide the correct answer in Chinese: If a cat's speed is 10 kilometers per hour, how long does it take for it to run 15 kilometers?" and "Based on the following question, provide the correct answer in English: If a cat's speed is 10 kilometers per hour, how long does it take for it to run 15 kilometers?".

[0067] It should be further explained that, based on any of the above embodiments, the training process of the above first large language model may include: using a multilingual data set to train the large language model to obtain an initial large language model; wherein the data in the multilingual data set includes news, encyclopedias, papers, and social media texts; using the task problem set to fine-tune the initial large language model to obtain the first large language model. This embodiment will use data from various fields to improve the coverage of knowledge, thereby improving the accuracy and comprehensiveness of the second large language model's reasoning.

[0068] S102, obtaining multiple-choice questions based on the initial reasoning results corresponding to each language and the multiple-choice question prompt template; wherein the multiple-choice question prompt template is a template including a logic for constructing the multiple-choice questions based on the initial reasoning results.

[0069] The multiple-choice prompt template in this embodiment is a template for constructing the logic of multiple-choice questions based on the initial reasoning results. This embodiment does not limit the specific method of constructing multiple-choice questions. For example, this embodiment can take the problem to be inferred as the question, and each initial reasoning result constitutes an option of the question; or this embodiment can combine the original question and the candidate answers in all languages ​​into a multiple-choice question format, and the reasoning process and answer order of each language correspond to each option to construct the final multiple-choice question. The multiple-choice prompt template in this embodiment may include question information about the options and each option, and each option is composed of the initial reasoning results. For example, the prompt word template of the multiple-choice question may be "{question} Please select the correct answer, option A (Chinese): {answers[0]}, option B (English): {answers[1]}". Assume that the candidate result list composed of the initial inference results corresponding to each language is answers=["The time it takes for the cat to run 15 kilometers is 1.5 hours", "The time it takes for the cat to run 15 kilometers is 1.5 hours"], then the constructed multiple-choice question is "If a cat's speed is 10 kilometers per hour, how long does it take to run 15 kilometers? Please choose the correct answer, option A (Chinese): The time it takes for the cat to run 15 kilometers is 1.5 hours, option B (English): The time it takes for the cat to run15 kilometers is 1.5 hours".

[0070] It should be further explained that, in order to improve the accuracy of multiple-choice question construction, the multiple-choice questions obtained based on the initial reasoning results corresponding to each language and the multiple-choice question prompt template may include:

[0071] S1021, obtain the inference process and initial inference result corresponding to each language based on the initial inference result corresponding to each language;

[0072] S1022, use the multiple-choice question objective, and construct a multiple-choice question with the inference process and initial inference result corresponding to each language as options.

[0073] In this embodiment, when constructing a multiple-choice question, not only the initial inference result is used, but also the inference process corresponding to each language is used, so that when using the second large language model for inference analysis, the inference process can be combined to improve the accuracy of multi-language context analysis.

[0074] S103, perform inference analysis using the second large language model based on the multiple-choice question to obtain the final inference result.

[0075] The second large language model in this embodiment is a model that can perform inference based on multiple-choice questions. The second large language model in this embodiment can be GPT; or it can also be Doubao, etc. The second large language model in this embodiment can be trained using a multi-language corpus and specifically trained with training data for multiple-choice question inference. When performing inference analysis in this embodiment, each language version can be analyzed independently first to generate the confidence level of each option (for example, the Chinese version supports option A with a confidence level of 85%). Compare the derivation results of each language, mark the conflict points, and conduct a specific analysis of the conflicting parts to obtain a more accurate answer. For example, the second large language model synthesizes the information of each language version to automatically select the final answer. For example, in the above example, whether it is Chinese or English, the final model will understand that 1.5 hours is the time required for the cat to run 15 kilometers, and the answers of each language version are the same. Therefore, the final answers are answers[0] and answers[1].

[0076] It should be further noted that, based on any of the above embodiments, the training process of the second large language model may include: obtaining a multilingual multiple-choice question dataset; wherein, the multilingual multiple-choice question dataset includes distraction options, and the distraction options include multimodal distraction data, and the multimodal distraction data includes mixed unit distraction and mixed semantic distraction; training the multilingual pre-trained model with the multilingual multiple-choice question dataset to obtain the second large language model. This embodiment takes into account that the grammar and word order differences of different languages may affect the accuracy of the model, so it uses distraction options for training. In multimodal machine learning tasks (such as combining data of different modalities such as images, texts, and voices), distraction options refer to artificially introduced or naturally existing distracting inputs, and the purpose is to disrupt the model's ability to understand or associate multimodal information. During the reasoning process, the data of different modalities are inconsistent in units, formats, or dimensions, resulting in difficulties for the model to align or fuse multimodal information. The mixed unit distraction in this embodiment refers to the cross-modal alignment problem caused by inconsistent data formats, units, or dimensions, mainly the interference at the physical level (units, formats, dimensions). The mixed semantic distraction in this embodiment refers to the logical understanding error caused by cross-modal semantic contradictions or irrelevancies. This embodiment trains the second large language model based on multimodal distraction data, enabling the model to reason more accurately.

[0077] It should be further noted that, based on any of the above embodiments, after using the second large language model to perform reasoning analysis based on multiple-choice questions and obtaining the final reasoning result, it may further include:

[0078] S1: Determine the target language corresponding to the question to be reasoned;

[0079] S2: Determine whether the final reasoning result includes the reasoning result corresponding to the target language;

[0080] S3: When the final reasoning result includes the reasoning result corresponding to the target language, send the reasoning result corresponding to the target language so that the language of the feedback reasoning result is consistent with the target language.

[0081] This embodiment will determine whether the final reasoning result contains the language version in the target language corresponding to the question to be reasoned. If so, the answer of the language version will be returned to the user. Otherwise, the final answer will be translated into the language of the question to be reasoned and then returned to the user. Assuming that the large language model selects the Chinese version and the English version of the answer from the multiple-choice question, since the user's original question is in Chinese, the system will directly return the Chinese version of the answer to the user. If the large language model only selects the English version of the answer from the multiple-choice question, the system will call the translation module at this time, translate the English answer into Chinese, and finally return it to the user. In practical applications, this translation mechanism can ensure that users can receive answers consistent with the language of their questions, while providing high-quality multilingual reasoning capabilities. This embodiment does not limit the specific method of translation. For example, this embodiment can determine the specific technical field of the problem to be reasoned based on the problem to be reasoned, thereby utilizing a dedicated translator in this field to improve the accuracy of the translation.

[0082] It should be further explained that, based on any of the above embodiments, after determining whether the final reasoning result includes the reasoning result corresponding to the target language, it may also include: when the final reasoning result does not include the reasoning result corresponding to the target language, translating the final reasoning result based on the target language using a translation model to obtain a translated reasoning result; and sending the translated reasoning result. This embodiment does not limit the specific translation model. For example, the translation model in this embodiment can be a cross-language reasoning result conversion method based on a logical dependency tree, wherein a logical dependency tree (LDT, Logical Dependency Tree) and a text synchronization conversion mechanism are constructed, and the predicate logical relationship of the original reasoning is retained during the translation process (such as , quantifier, → implied relationship); introduce cognitive symbol encoder (CSE), unify the encoding of text, mathematical symbols, and chart elements, support cross-language conversion of non-textual reasoning elements such as formulas and flowcharts, develop domain-aware routing network (DARN), call professional terminology library in real time, and build in 500+ vertical domain knowledge packages (medicine / law / engineering, etc.). The logic integrity verification module uses a dual-channel verification mechanism: Channel 1: original reasoning result → translation → back translation back to source language → logical equivalence verification; Channel 2: directly compare the similarity of the logical dependency tree structure before and after translation. In this process, a bilingual comparison logic report can be generated to mark the conversion process of key reasoning nodes. This model can improve the accuracy of translation. Alternatively, this embodiment can also construct a translation model corresponding to each field of the problem to be reasoned.

[0083] It should be further noted that, in order to improve the accuracy of the inference result, when the final inference result does not include the inference result corresponding to the target language, the final inference result is translated using a translation model based on the target language. After obtaining the translated inference result, it may further include: performing semantic detection on the translated inference result to determine whether the semantics is accurate; when it is accurate, sending the translated inference result; when it is inaccurate, performing inference analysis based on the final inference result and the translated inference result to obtain a corrected translated inference result. This embodiment performs semantic detection after obtaining the translated inference result, improving the accuracy of translation.

[0084] It should be further noted that, based on any of the above embodiments, after performing inference analysis using the second large language model based on a multiple-choice question to obtain the final inference result, it may further include: determining the confidence threshold corresponding to the final inference result; when the confidence threshold is lower than the set confidence threshold, sending a prompt message. This embodiment determines the confidence threshold corresponding to the final inference result, so that when the confidence threshold is low, the user can participate in manual inference.

[0085] An inference method provided by an embodiment of the present invention may include: S101, using a first large language model based on a target prompt template and a question to be inferred to obtain an initial inference result corresponding to each language; where the target prompt template is a template including the logic of the inference question; S102, obtaining a multiple-choice question based on the initial inference result corresponding to each language and a multiple-choice question prompt template; where the multiple-choice question prompt template is a template including the logic of constructing a multiple-choice question based on the initial inference result; S103, performing inference analysis using a second large language model based on the multiple-choice question to obtain the final inference result. Compared with the current inference process relying on a prompt word in a certain specific language, resulting in low inference accuracy, the present invention constructs a multiple-choice question based on the inference results corresponding to each language, enabling context analysis based on the inference results in multiple languages to obtain the final inference result. Since the final inference result can synthesize the inferences in multiple languages, the limitations of single-language model inference can be reduced, thus improving the accuracy of inference.

[0086] For the convenience of understanding the present invention, please specifically refer to Figure 2 , Figure 2 which is a flowchart example of a training method for a first large language model provided by an embodiment of the present invention, and may specifically include:

[0087] S201. Pre-train a large language model using a multilingual dataset to obtain a pre-trained inference model; where the multilingual dataset is a large-scale corpus including news, encyclopedias, and social media in various languages.

[0088] The pre-trained inference model in this embodiment is a model that can perform inference based on questions to obtain inference results. In this embodiment, the large language model is pre-trained on a multilingual dataset, which includes but is not limited to large-scale corpora such as news, encyclopedias, and social media texts in various languages. Through the training of the multilingual dataset, the model can master the basic grammar structures, common vocabulary, and semantic relationships between languages, providing support for cross-lingual inference and question understanding. As an alternative, existing open-source multilingual pre-trained models can also be directly used, such as directly using existing open-source multilingual pre-trained models like mBERT (Multilingual Model BERT), XLM-R (Cross-lingual Model - RoBERTa), etc.

[0089] S202. Fine-tune the pre-trained inference model using multilingual downstream task questions to obtain the first large language model.

[0090] This embodiment uses a dataset composed of multilingual downstream task questions to fine-tune the large language model. After completing the pre-training, the present invention further fine-tunes the large language model by using multilingual downstream task questions. Through this fine-tuning method, the present invention enables the model to not only be general in terms of language but also show high accuracy in specific downstream task fields.

[0091] The main purpose of the present invention is to provide a large language model inference method based on multilingual context, which creatively introduces self-generated multilingual context into the large language model.

[0092] Compared with the prior art, the advantage of the present invention is that during the inference process, the inference results of each language are concatenated into a multiple-choice question form, which can integrate the inference advantages of different languages and improve the inference accuracy and reliability in a multilingual environment. This method effectively solves the problems of difficult handling of language differences and inference results depending on a single input language existing in existing multilingual inference models, thereby improving the inference accuracy and stability of the model.

[0093] For the present invention to be more easily understood, specifically refer to Figure 3 , Figure 3 which is a flowchart example of an inference method provided by an embodiment of the present invention, and specifically may include:

[0094] S301. Obtain the question to be inferred, based on the target prompt template for each language, obtain the input information corresponding to each language, and combine the input information to obtain the complete input.

[0095] S302. Based on the complete input, use the first large language model to perform inference to obtain the inference process and initial inference results for each language.

[0096] In this embodiment, in order to obtain a multilingual context, the system will generate corresponding initial inference results for each language. For each language, candidate answers for the corresponding language are generated through the following steps: 1) For each language (such as Chinese, English, French, etc.), the system substitutes the string of the problem to be inferred in the original input into the target prompt templates of each language to obtain the complete input for the large model; 2) For each language, the model performs inference based on the domain knowledge (such as math problem sets, logical reasoning questions, etc.) trained in the fine-tuning stage to obtain the generation results for each target language, including the inference process and the inference result, etc.

[0097] S303, perform inference using the multiple-choice question prompt template based on the inference problem and the initial inference result for each language to obtain a multiple-choice question; wherein, the inference process and the inference result for each language are used as an option.

[0098] This embodiment splices the generated candidate answers (initial inference results) to obtain a multilingual context: the inference process and the initial inference result for each language are used as an option. In the form of a multiple-choice question, the system asks the large language model to select the most correct ones from multiple candidate answers. This selection does not depend solely on the inference result of a single language, but comprehensively considers the multilingual inference context.

[0099] S304, based on the multiple-choice question, use the second large language model to select the most correct inference result from multiple options;

[0100] S305, determine whether the language of the most correct inference result is the same as the language of the problem to be inferred.

[0101] S306, if they are the same, return the most correct inference result as the target inference result to the user.

[0102] S307, if they are not the same, translate the most correct inference result to make it consistent with the language of the problem to be inferred to obtain the target inference result, and return the target inference result to the user.

[0103] The technical solution of the present invention effectively combines the inference results of multiple language models through an innovative multi - language inference framework, thereby providing a cross - language inference solution. The main features of this technical solution include: (1) This technical solution can automatically identify the language of the user's question (the question to be inferred) and generate inference answers (inference results) in multiple language versions. This multi - language support enables users to obtain accurate inference results regardless of the language they use to ask questions. (2) The inference results in different languages are combined in the form of multiple - choice questions, and the large - language model synthesizes the information of each language version to select the final answer. This method can integrate the inference processes of different languages, thereby reducing the limitations of single - language model inference. (3) After selecting the final answer in the multiple - choice question, the system will determine whether the answer is consistent with the language of the user's original question. If not, the final answer will be translated back to the user's language to ensure that the user obtains an answer that matches the language of their question.

[0104] Traditional multi - language inference usually relies on translating the question into a single language for inference. The inference accuracy of this method is easily affected by the single - language ability of the model and cannot fully utilize the advantages of multi - language models. Many existing multi - language models only support basic translation or simple multi - language tasks and lack the integrated utilization of the cross - language inference ability of the model.

[0105] Compared with existing methods, the present invention has the following advantages:

[0106] 1. By simultaneously using the inference results in multiple language versions, the present invention not only enhances the robustness of inference but also effectively avoids the problem of error propagation in single - language inference.

[0107] 2. Using the inference results of multi - language inference to form multiple - choice questions allows the model to synthesize information from different languages and make a final choice. This approach significantly improves the inference accuracy compared with the existing technology. By integrating multi - language context information, the system can fully consider the differences of each language in inference and provide more comprehensive and multi - perspective answers. This multi - dimensional inference method has not been applied in the existing technology.

[0108] Next, the inference device provided by the embodiments of the present invention will be introduced. The inference device described below can be correspondingly referred to the inference method described above.

[0109] Figure 4 The structural schematic diagram of an inference device provided by an embodiment of the present invention may include:

[0110] An initial inference module 100, configured to use a first large - language model based on a target prompt template and a question to be inferred to obtain initial inference results corresponding to each language; wherein, the target prompt template is a template including the logic of the inference question;

[0111] A multiple-choice question determination module 200, configured to obtain a multiple-choice question based on the initial inference results corresponding to each language and a multiple-choice question prompt template, where the multiple-choice question prompt template is a template including a multiple-choice question logic constructed based on the initial inference results;

[0112] A final inference module 300, configured to perform inference analysis on the multiple-choice question using a second large language model to obtain a final inference result.

[0113] Further, based on the above embodiments, the initial inference module 100 may include:

[0114] A target language determination module, configured to obtain the question to be inferred and the target language corresponding to the question to be inferred;

[0115] A first input information determination unit, configured to obtain first input information based on the question to be inferred and the target language using a target language prompt template;

[0116] A second input information determination unit, configured to obtain second input information based on the question to be inferred and the remaining languages using a language prompt unit corresponding to the remaining languages; the target prompt template includes the target language prompt template and the language prompt template;

[0117] An initial inference result determination unit, configured to obtain the initial inference results corresponding to each language based on the first input information and the second input information using the first large language model.

[0118] Further, based on any of the above embodiments, the multiple-choice question determination module 200 may include:

[0119] An inference process and result determination unit, configured to obtain the inference process and the initial inference results corresponding to each language based on the initial inference results corresponding to each language;

[0120] A multiple-choice question construction unit, configured to use the multiple-choice question target and use the inference process and the initial inference results corresponding to each language as options to construct the multiple-choice question.

[0121] Further, based on the above embodiments, the above inference device may further include:

[0122] A target language determination module, configured to determine the target language corresponding to the question to be inferred;

[0123] A judgment module, configured to determine whether the final inference result includes an inference result corresponding to the target language;

[0124] Inference result range module, configured to send the inference result corresponding to the target language when the final inference result includes the inference result corresponding to the target language, so that the language of the feedback inference result is consistent with the target language.

[0125] Further, based on the above embodiment, the above-mentioned inference device may further include:

[0126] Translation module, configured to translate the final inference result using a translation model based on the target language to obtain a translated inference result when the final inference result does not include the inference result corresponding to the target language;

[0127] Sending module, configured to send the translated inference result.

[0128] Further, based on the above embodiment, the above-mentioned inference device may further include:

[0129] Semantic detection module, configured to perform semantic detection on the translated inference result to determine whether the semantics is accurate;

[0130] Translated inference result sending module, configured to send the translated inference result when it is accurate;

[0131] Correction module, configured to perform inference analysis based on the final inference result and the translated inference result to obtain a corrected translated inference result when it is inaccurate.

[0132] Further, based on the above embodiment, the above-mentioned inference device may further include:

[0133] Confidence threshold determination module, configured to determine the confidence threshold corresponding to the final inference result;

[0134] Prompt information sending module, configured to send prompt information when the confidence threshold is lower than the set confidence threshold.

[0135] Further, based on the above embodiment, the above-mentioned inference device may further include:

[0136] Initial large language model determination module, configured to train a large language model using a multilingual dataset to obtain an initial large language model; wherein, the data in the multilingual dataset includes news, encyclopedias, papers, and social media texts;

[0137] Second large language model training module, configured to fine-tune the initial large language model using a task question set to obtain the first large language model.

[0138] Further, based on the above embodiment, the above-mentioned inference device may further include:

[0139] A multilingual multiple-choice question dataset determination module for obtaining a multilingual multiple-choice question dataset; wherein, the multilingual multiple-choice question dataset includes distractors, and the distractors include multimodal interference data, and the multimodal interference data includes mixed unit interference and mixed semantic interference;

[0140] A second large language model training module for training a multilingual pre-trained model using the multilingual multiple-choice question dataset to obtain the second large language model.

[0141] Further, based on any of the above embodiments, the multiple-choice question prompt template includes question information about the options and each option, and each option is composed of the initial inference result.

[0142] Further, based on any of the above embodiments, the question to be inferred includes at least one of an intelligent education question and a medical diagnosis question.

[0143] It should be noted that the modules and units in the above inference device can be changed in order before and after without affecting the logic.

[0144] Figure 4 The description of the features in the corresponding embodiments can be referred to Figure 4 the relevant descriptions of the corresponding embodiments, which will not be elaborated here one by one.

[0145] The inference device provided by the embodiments of the present invention may include: an initial inference module 100 for obtaining an initial inference result corresponding to each language based on a target prompt template and a question to be inferred using a first large language model; wherein, the target prompt template is a template including the logic of the inference question; a multiple-choice question determination module 200 for obtaining multiple-choice questions based on the initial inference results corresponding to each language and a multiple-choice question prompt template; wherein, the multiple-choice question prompt template is a template including the logic of constructing multiple-choice questions based on the initial inference results; a final inference module 300 for performing inference analysis using a second large language model based on the multiple-choice questions to obtain a final inference result. Compared with the current inference process relying on a prompt word in a specific language, resulting in low inference accuracy, the present invention constructs multiple-choice questions based on the inference results corresponding to each language, so that context analysis can be performed based on the multilingual inference results to obtain a final inference result. Since the final inference result can integrate the inferences of multiple languages, the limitations of single-language model inference can be reduced, so the inference accuracy can be improved.

[0146] Next, an inference device provided by the embodiments of the present invention will be introduced, and the inference device described below can be mutually referred to the inference method described above.

[0147] Figure 5The following is a schematic structural diagram of an inference device provided by an embodiment of the present invention. As Figure 5 shown, the inference device includes: a memory 60 for storing computer programs;

[0148] a processor 61 for implementing the steps of the inference method in the above embodiments when executing the computer program.

[0149] The inference device provided in this embodiment may include, but is not limited to, a smart phone, a tablet computer, a notebook computer, a desktop computer, etc.

[0150] Among them, the processor 61 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 61 may be implemented in at least one of the following hardware forms: digital signal processing (DSP), field-programmable gate array (FPGA), and programmable logic array (PLA). The processor 61 may also include a main processor and a coprocessor. The main processor is a processor for processing data in the wake state, also known as the central processing unit (CPU); the coprocessor is a low-power processor for processing data in the standby state. In some embodiments, the processor 61 may be integrated with a graphics processing unit (GPU), and the GPU is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 61 may further include an artificial intelligence (AI) processor for processing computational operations related to machine learning.

[0151] The memory 60 may include one or more computer-readable storage media, and the computer-readable storage media may be non-transitory. The memory 60 may further include a high-speed random access memory and a non-volatile memory, such as one or more disk storage devices and flash storage devices. In this embodiment, the memory 60 is at least used to store the following computer program 601. After the computer program is loaded and executed by the processor 61, it can implement the relevant steps of the inference method disclosed in any of the foregoing embodiments. In addition, the resources stored in the memory 60 may further include an operating system 602 and data 603, etc., and the storage method may be temporary storage or permanent storage. Among them, the operating system 602 may include Windows, Unix, Linux, etc. The data 603 may include, but is not limited to, the data required for the inference method.

[0152] In some embodiments, the inference device may further include a display screen 62, an input / output interface 63, a communication interface 64, a power supply 65, and a communication bus 66.

[0153] Those skilled in the art can understand that Figure 5 the structure shown in does not constitute a limitation on the inference device, and it may include more or fewer components than shown in the figure.

[0154] It can be understood that if the inference method in the above embodiments is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the current technology, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and executes all or part of the steps of the methods in the various embodiments of the present invention. The aforementioned storage media include: USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), electrically erasable programmable ROMs, registers, hard disks, removable disks, CD-ROMs, magnetic disks, or optical disks, etc., all of which can store program codes.

[0155] Based on this, the embodiments of the present invention further provide a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the inference method as described above.

[0156] The above has introduced in detail an inference method provided by the embodiments of the present invention. The various embodiments in the specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0157] Those skilled in the art can further realize that the units and algorithm steps of each example described in combination with the embodiments disclosed in this article can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described according to functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0158] The above has introduced in detail a reasoning method, device, equipment, readable storage medium and program product provided by the present invention. Specific examples are used in this article to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and modifications can still be made to the present invention, and these improvements and modifications also fall within the protection scope of the present invention.

Claims

1. A reasoning method, characterized in that: include: Based on the target prompt template and the question to be inferred, the first language model is used to obtain the initial inference results corresponding to each language; wherein the target prompt template is a template including the inference question logic; Based on the initial reasoning results and multiple-choice question prompt templates corresponding to each language, a multiple-choice question is obtained; wherein the multiple-choice question prompt template is a template including a multiple-choice question logic constructed based on the initial reasoning results; Based on the multiple-choice question, the second largest language model is used to perform reasoning analysis to obtain a final reasoning result.

2. The inference method according to claim 1, characterized in that: Based on the target prompt template and the question to be inferred, the first language model is used to obtain the initial inference results corresponding to each language, including: Obtaining the problem to be inferred and the target language corresponding to the problem to be inferred; Obtaining first input information based on the problem to be inferred and the target language using a target language prompt template; Based on the problem to be inferred and the remaining language, the second input information is obtained by using the language prompt template corresponding to the remaining language; the target prompt template includes the target language prompt template and the language prompt template; The initial inference results corresponding to each language are obtained by using the first large language model based on the first input information and the second input information.

3. The inference method according to claim 1, characterized in that: Based on the initial reasoning results and multiple-choice question prompt templates corresponding to each language, multiple-choice questions are obtained, including: Based on the initial reasoning results corresponding to the respective languages, obtaining the reasoning process and the initial reasoning results corresponding to the respective languages; The multiple-choice question is constructed by using the multiple-choice question target and taking the reasoning process and the initial reasoning result corresponding to each language as options.

4. The inference method according to any one of claims 1 to 3, characterized in that: After performing reasoning analysis based on the multiple-choice question using the second language model to obtain a final reasoning result, the method further includes: Determining a target language corresponding to the problem to be reasoned; determining whether the final inference result includes an inference result corresponding to the target language; When the final inference result includes an inference result corresponding to the target language, the inference result corresponding to the target language is sent so that the language of the fed-back inference result is consistent with the target language.

5. The inference method according to claim 4, characterized in that: After determining whether the final inference result includes an inference result corresponding to the target language, the method further includes: When the final reasoning result does not include a reasoning result corresponding to the target language, translating the final reasoning result using a translation model based on the target language to obtain a translated reasoning result; The inference result of the translation is sent.

6. The inference method according to claim 5, characterized in that: When the final reasoning result does not include a reasoning result corresponding to the target language, the final reasoning result is translated based on the target language using a translation model to obtain a translated reasoning result, further comprising: Performing semantic detection on the inference result of the translation to determine whether the semantics are accurate; When accurate, sending the inference result of the translation; When it is inaccurate, reasoning analysis is performed based on the final reasoning result and the translated reasoning result to obtain a corrected translated reasoning result.

7. The inference method according to claim 1, characterized in that: After performing reasoning analysis based on the multiple-choice question using the second language model to obtain a final reasoning result, the method further includes: Determining a confidence threshold corresponding to the final reasoning result; When the confidence threshold is lower than the set confidence threshold, a prompt message is sent.

8. The inference method according to claim 1, characterized in that: The training process of the first language model includes: Using a multilingual dataset to train a large language model to obtain an initial large language model; wherein the data in the multilingual dataset includes news, encyclopedias, papers, and social media texts; The initial large language model is fine-tuned using the task problem set to obtain the first large language model.

9. The inference method according to claim 1, characterized in that: The training process of the second largest language model includes: Acquire a multilingual multiple-choice question dataset; wherein the multilingual multiple-choice question dataset includes interference options, the interference options include multimodal interference data, and the multimodal interference data includes mixed unit interference and mixed semantic interference; The multilingual multiple-choice question dataset is used to train a multilingual pre-trained model to obtain the second largest language model.

10. The inference method according to claim 1, characterized in that: The multiple-choice question prompt template includes question information about options and each option, and each option is composed of the initial reasoning result.

11. The inference method according to claim 1, characterized in that: The problem to be inferred includes at least one of an intelligent education problem and a medical diagnosis problem.

12. An inference device, characterized in that: include: An initial reasoning module, used to obtain initial reasoning results corresponding to each language based on a target prompt template and a question to be reasoned using a first language model; wherein the target prompt template is a template including reasoning question logic; A multiple-choice question determination module, used to obtain multiple-choice questions based on the initial reasoning results and multiple-choice question prompt templates corresponding to each language; wherein the multiple-choice question prompt template is a template including a multiple-choice question logic constructed based on the initial reasoning results; The final reasoning module is used to perform reasoning analysis based on the multiple-choice question using the second largest language model to obtain a final reasoning result.

13. An inference device, characterized in that: include: Memory for storing computer programs; A processor, configured to execute the computer program to implement the steps of the reasoning method according to any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the reasoning method according to any one of claims 1 to 11 are implemented.

15. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the reasoning method according to any one of claims 1 to 11 are implemented.

Images