Method, system, computing device and electronic device for processing interaction information

Abstract

The application discloses a kind of interactive information processing method, system, computing device and electronic equipment, it is related to big model technology, natural language processing field.Therein, the method can include: obtaining input information, and inputting input information to dialogue model for analysis, obtain at least one candidate reply information;At least one candidate reasoning step of candidate reply information is obtained, wherein candidate reasoning step is used to indicate the process that dialogue model analyzes input information;Candidate reasoning step and corresponding reference reasoning step are semantically aligned, to obtain alignment result, wherein reference reasoning step is used to indicate the process that dialogue model analyzes reference input information, and the language type of reference input information is different from the language type of input information;Based on alignment result, from at least one candidate reply information, determine the target reply information that matches with input information.The application solves the technical problem that dialogue model has big limitation to interactive information processing.

Description

Technical Field

[0001] This application relates to the fields of large model technology and natural language processing, and more specifically, to a method, system, computing device, and electronic device for processing interactive information. Background Technology

[0002] Currently, in the field of natural language processing, especially in the development of dialogue models, the imbalance of resource language types has become a significant challenge. This imbalance is mainly reflected in the fact that some languages ​​(high-resource languages, such as English and Chinese) possess a large amount of high-quality training data, while other languages ​​(low-resource languages) have relatively scarce data.

[0003] The differences in the distribution of training data for the aforementioned languages ​​directly lead to significant performance differences in dialogue models across different language environments. This is mainly manifested in the following ways: high-resource languages ​​show a clear advantage. For example, after deep training, dialogue models perform exceptionally well in high-resource languages ​​such as English and Chinese, accurately understanding user intent and generating coherent and reasonable responses. Low-resource languages, on the other hand, perform poorly. For example, even with supervised fine-tuning techniques, dialogue models struggle to achieve the same level of performance for languages ​​with limited data, especially when dealing with complex reasoning or understanding specific cultural contexts, where their performance is often unsatisfactory.

[0004] The aforementioned performance gaps not only limit the application of dialogue models in multilingual scenarios but also hinder effective communication among users worldwide. Particularly in cross-language scenarios, dialogue models cannot accurately transfer their reasoning abilities, resulting in a lack of consistency across different languages ​​when processing language-independent knowledge, such as mathematical calculations, logical analysis, and code interpretation. This impacts their universality and cross-cultural adaptability. Therefore, the technical challenge of dialogue models' significant limitations in processing interactive information remains.

[0005] There is currently no effective solution to the above problems. Summary of the Invention

[0006] This application provides a method, system, computing device, and electronic device for processing interactive information, in order to at least solve the technical problem that dialogue models have significant limitations in processing interactive information.

[0007] According to one aspect of the embodiments of this application, a method for processing interactive information is provided. The method may include: acquiring input information and inputting the input information into a dialogue model for analysis to obtain at least one candidate response; acquiring at least one candidate inference step from the candidate response, wherein the candidate inference step represents the process by which the dialogue model analyzes the input information; semantically aligning the candidate inference step with a corresponding reference inference step to obtain an alignment result, wherein the reference inference step represents the process by which the dialogue model analyzes reference input information, and the language type of the reference input information is different from the language type of the input information; and based on the alignment result, determining a target response that matches the input information from the at least one candidate response.

[0008] According to another aspect of the embodiments of this application, another method for processing interactive information is provided. The method may include: acquiring query information in a target language type, wherein the query information represents information to be answered in a dialogue scenario; inputting the query information into a dialogue model for analysis to obtain at least one candidate response, wherein the dialogue model is trained based at least on information samples in the target language type and information samples in the source language type, and the sample size of the information samples in the target language type is lower than the sample size of the information samples in the source language type; acquiring at least one candidate inference step of the candidate response, wherein the candidate inference step represents the process by which the dialogue model analyzes the input information; semantically aligning the candidate inference step with the corresponding reference inference step to obtain an alignment result, wherein the reference inference step represents the process by which the dialogue model analyzes reference input information in the source language type, and the language type of the reference inference information is different from the language type of the query information; and based on the alignment result, determining target response information that matches the query information from at least one candidate response.

[0009] According to another aspect of the embodiments of this application, another method for processing interactive information is provided. The method may include: responding to an input operation on an interface and displaying input information on the interface; responding to an information response operation on the interface and displaying target response information matching the input information on the interface, wherein the target response information is determined from at least one candidate response information based on alignment results, the alignment results being obtained by semantically aligning at least one candidate inference step with its corresponding reference inference step, the candidate inference step representing the process of the dialogue model analyzing the input information, the reference inference step representing the process of the dialogue model analyzing reference input information, the language type of the reference input information being different from the language type of the input information, and the candidate response information being obtained by the dialogue model from analyzing the input information.

[0010] According to another aspect of the embodiments of this application, a method for processing interactive information is provided. The method may include: obtaining input information by calling a first interface, wherein the first interface includes a first parameter, and the parameter value of the first parameter includes the input information; inputting the input information into a dialogue model for analysis to obtain at least one candidate response; obtaining at least one candidate inference step from the candidate response information, wherein the candidate inference step represents the process by which the dialogue model analyzes the input information; semantically aligning the candidate inference step with a corresponding reference inference step to obtain an alignment result, wherein the reference inference step represents the process by which the dialogue model analyzes reference input information, and the language type of the reference input information is different from the language type of the input information; based on the alignment result, determining a target response that matches the input information from the at least one candidate response; and outputting the target response information by calling a second interface, wherein the second interface includes a second parameter, and the parameter value of the second parameter includes the target response information.

[0011] According to another aspect of this embodiment of the application, an interactive information processing system is provided. The system may include: an information input terminal for acquiring input information; an information processing terminal for inputting the input information into a dialogue model for analysis to obtain at least one candidate response; acquiring at least one candidate inference step of the candidate response, wherein the candidate inference step represents the process by which the dialogue model analyzes the input information; semantically aligning the candidate inference step with a corresponding reference inference step to obtain an alignment result, wherein the reference inference step represents the process by which the dialogue model analyzes reference input information, and the language type of the reference input information is different from the language type of the input information; based on the alignment result, determining a target response that matches the input information from at least one candidate response; and an information output terminal for outputting the target response.

[0012] According to another aspect of the embodiments of this application, a computing device is also provided, including: a memory storing an executable program; and a processor for running the program, wherein the program executes the methods in various embodiments of this application when it runs.

[0013] According to another aspect of the embodiments of this application, an electronic device is also provided, including: a memory storing an executable program; and a processor connected to the memory via a bus for running the program, wherein the program executes the methods in various embodiments of this application when it runs.

[0014] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored executable program, wherein, when the executable program is running, it controls the device where the computer-readable storage medium is located to perform the methods of various embodiments of this application.

[0015] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program that, when executed by a processor, implements the methods of various embodiments of this application.

[0016] According to another aspect of the embodiments of this application, a computer program product is also provided, including a non-volatile computer-readable storage medium storing a computer program, which, when executed by a processor, implements the methods in various embodiments of this application.

[0017] According to another aspect of the embodiments of this application, a computer program is also provided, which, when executed by a processor, implements the methods of the various embodiments of this application.

[0018] In this embodiment, if a dialogue model is needed to respond to input information, the acquired input information can be input into the dialogue model, and the dialogue model can analyze the input information to obtain candidate response information. At least one candidate inference step corresponding to the candidate response information can be obtained; that is, the process of analyzing the input information using the dialogue model to obtain at least one candidate response information can be obtained. Reference inference steps obtained by the dialogue model analyzing reference input information of a different language type can be obtained, and the candidate inference steps and reference inference steps can be semantically aligned to obtain an alignment result. Based on the alignment result, the target response information capable of responding to the input information can be determined from at least one candidate response information.

[0019] In this embodiment, if the input information is in a low-resource language, semantic alignment with the reference reasoning step allows for the transfer of mature reasoning capabilities from high-resource languages ​​to low-resource languages, significantly improving the versatility and adaptability of the dialogue model in multilingual environments. Furthermore, reasoning step alignment ensures consistency in the logical reasoning process across different languages, avoiding semantic errors or information distortion caused by language differences. In summary, this application, through reasoning step alignment and semantic alignment mechanisms, successfully improves the reasoning ability and output quality of dialogue models in cross-language scenarios, thereby reducing the limitations of dialogue models in processing interactive information and solving the problem of significant limitations in interactive information processing.

[0020] It is worth noting that the general description above and the detailed description that follow are merely for illustrative purposes and do not constitute a limitation on this application. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0022] Figure 1 This is a schematic diagram illustrating an application scenario of an interactive information processing method according to an embodiment of this application;

[0023] Figure 2 This is a flowchart of an interactive information processing method according to an embodiment of this application;

[0024] Figure 3 This is a flowchart of another method for processing interactive information according to an embodiment of this application;

[0025] Figure 4 This is a flowchart of another method for processing interactive information according to an embodiment of this application;

[0026] Figure 5 This is a flowchart of another method for processing interactive information according to an embodiment of this application;

[0027] Figure 6 This is a schematic diagram of an interactive information processing system according to an embodiment of this application;

[0028] Figure 7 This is a schematic diagram of a common capability in related technologies: cross-language alignment.

[0029] Figure 8(a) is a schematic diagram of a literal translation method in a related technology;

[0030] Figure 8(b) is a schematic diagram of a translation-generation method in a related technology;

[0031] Figure 9 This is a schematic diagram of a cross-language ability transfer process based on inference process alignment according to an embodiment of this application;

[0032] Figure 10 This is a schematic diagram of an interactive information processing device according to an embodiment of this application;

[0033] Figure 11 This is a schematic diagram of another interactive information processing device according to an embodiment of this application;

[0034] Figure 12 This is a schematic diagram of another interactive information processing device according to an embodiment of this application;

[0035] Figure 13 This is a schematic diagram of another interactive information processing device according to an embodiment of this application;

[0036] Figure 14 This is a structural block diagram of a computing device according to an embodiment of this application;

[0037] Figure 15 This is a structural block diagram of an electronic device according to an embodiment of this application. Detailed Implementation

[0038] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0039] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0040] The technical solution provided in this application is mainly implemented using large-scale model technology. Here, "large-scale model" refers to a deep learning model with a large number of parameters, typically containing hundreds of millions, tens of billions, hundreds of billions, trillions, or even tens of trillions of parameters. Large-scale models can also be called foundation models. They are pre-trained using large-scale unlabeled corpora to produce pre-trained models with hundreds of millions of parameters. Such models can adapt to a wide range of downstream tasks and have good generalization ability. Examples include Large Language Models (LLMs) and multi-modal pre-training models.

[0041] It should be noted that, in practical applications, large models can be fine-tuned using a small number of samples to adapt them to different tasks. For example, large models can be widely applied in Natural Language Processing (NLP), computer vision, and speech processing. Specifically, they can be applied to computer vision tasks such as Visual Question Answering (VQA), Image Captioning (IC), and Image Generation, as well as NLP tasks such as text-based sentiment classification, text summarization, and machine translation. Therefore, the main application scenarios for large models include, but are not limited to, digital assistants, intelligent robots, search, online education, office software, e-commerce, and intelligent design. In this embodiment, a dialogue model is used to analyze the input information to obtain and explain the target response information that has been addressed to the input information.

[0042] First, some nouns or terms that appear in the description of the embodiments of this application shall be interpreted as follows:

[0043] Large Language Models (LLMs) are large-scale deep neural networks that learn from massive amounts of unsupervised text.

[0044] Multilingual Large Language Model (MLLM) is a large-scale deep neural network that can understand and generate multiple languages ​​and learn from massive amounts of multilingual text.

[0045] Supervised fine-tuning (SFT) is a technique for training large models to follow instructions and align with human preferences.

[0046] Bidirectional Encoder Representations from Transformers (BERT) is a natural language processing model that uses a transformer architecture. Its key feature is that it can understand that the meaning of a word in a sentence depends on the information of the words before and after it, rather than relying solely on the preceding or following context.

[0047] Language-agnostic BERT Sentence Embedding (LaBSE) is a cross-language sentence embedding model based on BERT that maps sentences from different languages ​​to the same vector space.

[0048] General competence refers to the ability of a large model to understand and process the same concepts and knowledge expressed in different languages, and the nature of this knowledge (such as common sense, mathematics, code, etc.) is language-independent.

[0049] According to an embodiment of this application, a method for processing interactive information is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0050] Considering the large number of model parameters in large models and the limited computing resources of mobile terminals, the interactive information processing method provided in this application embodiment can be applied to, for example, Figure 1 The application scenarios shown are not limited to these. In, for example... Figure 1 In the application scenario shown, the large model is deployed on server 30, which can be in the cloud. Server 30 can connect to one or more terminal devices 10 via LAN, WAN, the Internet, or other types of data networks. Terminal devices 10 can be client devices, including but not limited to smartphones, tablets, laptops, PDAs, personal computers, smart home devices, and in-vehicle devices. These client devices collectively constitute the client relative to the server. A graphical user interface (GUI) for obtaining input information can be deployed on the client device. This GUI can be an interface on a dialogue system that allows interaction with the dialogue model. Terminal device 10 can interact with the user through the GUI to invoke the large model, thereby implementing the interactive information processing method provided in this embodiment. Server 30 and terminal device 10 can exchange information via network 20. Database 40 in the server can deploy a large model, such as a dialogue model. In this embodiment, database 40 stores a target model for analyzing input information to obtain candidate response information. This model can be invoked through server 30 to analyze input information and obtain candidate response information.

[0051] In this embodiment, the system comprising terminal device 10, network 20, and server 30 can perform the following steps: When a user needs to respond to input information, they can perform input or selection operations on the operation interface of the corresponding terminal device 10 to upload the input information to the terminal device 10. Terminal device 10 can send the input information to server 30 via network 20. After receiving the input information, server 30 can perform the following steps: Step S102, obtain the input information and input it into the dialogue model for analysis to obtain at least one candidate response information; Step S104, obtain at least one candidate inference step of the candidate response information; Step S106, perform semantic alignment between the candidate inference step and the corresponding reference inference step to obtain an alignment result; Step S108, based on the alignment result, determine the target response information that matches the input information from the at least one candidate response information. The target response information generated by the dialogue model can be output to terminal device 10 via network 20 to be displayed on the operation interface of the terminal device for user viewing.

[0052] In this embodiment, if the input information is in a low-resource language, semantic alignment with the reference reasoning step allows for the transfer of mature reasoning capabilities from high-resource languages ​​to low-resource languages, significantly improving the versatility and adaptability of the dialogue model in multilingual environments. Furthermore, reasoning step alignment ensures consistency in the logical reasoning process across different languages, avoiding semantic errors or information distortion caused by language differences. In summary, through reasoning step alignment and semantic alignment mechanisms, the reasoning ability and output quality of the dialogue model are successfully improved in cross-language scenarios, thereby reducing the limitations of the dialogue model in processing interactive information and solving the problem of significant limitations in interactive information processing.

[0053] It should be noted that with the rapid development of high-performance computing units, the methods provided in this application embodiment can also be applied to model-in-machine systems in other application scenarios. In one optional embodiment, the model-in-machine system has multiple built-in models, and users can select one model to adjust as needed to obtain their own model. The high-performance computing unit built into the model-in-machine system can then directly call the adjusted model to execute the methods provided in this application embodiment. In another optional embodiment, the large model-in-machine system has a pre-trained model built-in, and the high-performance computing unit built into the model-in-machine system can then directly call that model to execute the methods provided in this application embodiment.

[0054] Furthermore, when users need to train their own models, they can upload their own datasets via the client. These datasets are then sent to the server, allowing the server to adjust the pre-trained model using the dataset to obtain the user's customized model, which can then be deployed to the production environment. To facilitate users' model adjustment needs, the server provides complete adjustment tools, development frameworks, and processes, supporting multiple adjustment strategies. This allows the adjusted model to better adapt to different application domains and achieve a high degree of customization.

[0055] Under the aforementioned operating environment, this application provides the following: Figure 2 The method for processing interactive information shown is illustrated. It should be noted that the method for processing interactive information in this embodiment can be derived from... Figure 1 The mobile terminal of the illustrated embodiment is executed. Figure 2 This is a flowchart of an interactive information processing method according to an embodiment of this application, such as... Figure 2 As shown, the method may include the following steps:

[0056] Step S202: Obtain input information and input the input information into the dialogue model for analysis to obtain at least one candidate response.

[0057] In the technical solution provided by step S202 of this application, the input information can be natural language text provided by the user, intended to obtain a response or answer through the dialogue model. In the embodiments of this application, the input information can be questions, statements, instructions, or requests made in different languages. The input information can be used to trigger the entire process of the dialogue model understanding the questions in the input information and generating corresponding responses. In the cross-linguistic ability transfer method based on reasoning process alignment, the diversity and complexity of the input information require the model to generalize to expressions in different languages ​​and understand the intent behind the input information and the knowledge domains involved. The input information can be a target language question, for example, a multilingual question (query). The input information can come from different language types, such as Japanese, Spanish, Russian, etc., reflecting the user's needs in different language scenarios.

[0058] Optionally, the dialogue model can be a deep learning model capable of understanding and generating natural language dialogue. It is a core component in this embodiment, used to analyze input information and generate candidate response information. The dialogue model can be a large model based on the Transformer architecture. The dialogue model can be trained on a large scale, either unsupervised or supervised, and possess the ability to process multilingual text. After receiving input information, the dialogue model can perform deep semantic analysis to understand the intent and background knowledge of the input information. It can utilize internal components such as word embeddings, attention mechanisms, and decoding to parse the structure and semantics of the input information, and then generate one or more candidate response information based on the patterns and rules learned by the dialogue model. The generation capability of the dialogue model is key to achieving cross-language reasoning alignment, enabling the generation of high-quality, semantically accurate, and structurally sound candidate response information, providing a foundation for subsequent reasoning alignment. Optionally, the dialogue model can be a large multilingual model.

[0059] Optionally, candidate response information consists of multiple possible reply options generated by the dialogue model based on the input information. Unlike a single response, generating multiple candidate responses increases the model's flexibility in handling complex and ambiguous inputs. Through a subsequent filtering process, the dialogue model can select a suitable response from the candidate responses as the target response. Candidate response information can be candidate answers.

[0060] In this embodiment, if a response is required to the input information, the acquired input information can be input into the dialogue model, and the dialogue model can be used to perform in-depth analysis on the input information to obtain at least one candidate response information.

[0061] Optionally, if user input is received, the input information can be fed into a pre-trained dialogue model. The dialogue model is a deep neural network, pre-trained on large-scale text data, possessing powerful language understanding and generation capabilities. After receiving input information, the dialogue model can perform internal processing on the text, including word embedding, semantic parsing, and contextual understanding.

[0062] Optionally, within the dialogue model, the input information is decomposed and understood. Based on pre-learned knowledge and patterns, the dialogue model performs in-depth analysis of the questions posed in the input information. During the analysis, the semantics and intent of the input text can be understood, and possible responses can be inferred from the input information to generate at least one candidate response.

[0063] Optionally, the series of candidate responses obtained by analyzing the input information are semantically related to the input information, but their specific forms (e.g., content, structure, and expression) can be diverse. The purpose of generating multiple candidate responses is to conduct more refined screening and evaluation in subsequent steps, thereby selecting a more context-appropriate and accurate response.

[0064] In this embodiment, the above process not only generates responses to specific questions in the input information, but also provides multiple alternative responses (candidate response information) through a multi-candidate response information generation strategy, increasing the flexibility and adaptability of the dialogue model's responses. These steps provide rich material for subsequent advanced processing such as semantic alignment in the reasoning process. In other words, by generating diverse candidate response information, a solid foundation is laid for subsequent semantic alignment and high-quality response selection.

[0065] Step S204: Obtain at least one candidate reasoning step of candidate response information.

[0066] In the technical solution provided in step S204 of this application, candidate inference steps can be used to represent the process by which the dialogue model analyzes input information. Multiple candidate inference steps can also be referred to as a sequence of inference steps (Steps). A candidate inference step refers to a sentence or paragraph extracted from candidate response information that can independently express a complete inference unit. Candidate inference steps are an explicit representation of the internal inference process of the dialogue model when understanding input information and generating responses. Each candidate inference step carries a part of logical reasoning, which can be a statement of facts, rule application, logical reasoning, or conclusion derivation, etc.

[0067] In this embodiment, after analyzing the input information using a dialogue model to obtain candidate response information, at least one candidate reasoning step can be obtained from the candidate response information.

[0068] Optionally, by decomposing candidate response information into a series of candidate reasoning steps, the reasoning process originally implicit in the dialogue model can be made explicit, making the reasoning path clearer and more analyzable. In cross-language ability transfer, candidate reasoning steps, as finer-grained alignment units, help to find logically identical or similar reasoning units in responses from different languages, thereby evaluating the cross-language consistency of reasoning steps. The logical rationality and semantic accuracy of each candidate reasoning step can be evaluated individually, providing a foundation for answer selection based on reasoning consistency in subsequent steps.

[0069] Optionally, obtaining at least one candidate reasoning step for candidate response information is a key technology for multilingual large models to handle complex problems. The purpose is to refine the model's reasoning process to facilitate subsequent semantic alignment and quality screening.

[0070] Optionally, decomposing candidate response information into a series of reasoning steps helps to understand the model's decision-making process more deeply and provides material for subsequent screening based on the consistency of the reasoning process. By making the reasoning steps explicit, it is easier to find logically identical or similar reasoning units in responses from different languages, thereby assessing the cross-language consistency of the reasoning steps. The logical rationality and semantic accuracy of each reasoning step can be evaluated individually, helping to control the overall quality of candidate response information and ensuring the accuracy of the final response in multiple language environments.

[0071] For example, using NLP techniques such as syntactic analysis, semantic role labeling, and discourse analysis, the logical structure and semantic relationships within candidate responses can be broken down into independent reasoning steps. For each reasoning step, a corresponding representation can be found in multiple languages, and this process can be implemented through machine translation.

[0072] In this embodiment, the above method effectively refines the decision-making process of the multilingual large model, transforming abstract reasoning into analyzable and comparable concrete steps. By making the model's internal logic explicit, it not only enhances the understanding and control of the dialogue model's multilingual processing capabilities but also provides a solid foundation for subsequent reasoning consistency assessment and cross-language capability transfer. This method significantly improves the accuracy and logical coherence of responses in multilingual scenarios, ensuring consistency in knowledge representation and reasoning paths across different languages. This promotes the realization of high-quality multilingual services, especially in resource-scarce language environments, effectively improving the dialogue model's versatility and user satisfaction.

[0073] Step S206: Semantically align the candidate reasoning steps with the corresponding reference reasoning steps to obtain the alignment results.

[0074] In the technical solution provided by step S206 of this application, the reference reasoning step can be used to represent the process by which the dialogue model analyzes reference input information. The language type of the reference input information is different from the language type of the input information.

[0075] Optionally, the reference reasoning step can be each reasoning step in the source language, such as each reasoning step in a Chinese / English response. A reference reasoning step refers to a sequence of reasoning steps generated by a dialogue model based on input information in a high-resource language (such as Chinese or English). This sequence of reasoning steps can be considered a standard or reference for comparison and alignment with candidate reasoning steps in a low-resource language to ensure cross-language consistency in the reasoning process. The reference reasoning step serves as a benchmark for semantic alignment, providing clear logical paths and semantic expressions, enabling the evaluation and optimization of candidate reasoning steps in a low-resource language based on known standards.

[0076] Optionally, the alignment result can refer to a quantitative indicator obtained by calculating the semantic similarity between the candidate inference step and the reference inference step, usually expressed as cosine similarity. The alignment result can also be called semantic alignment, used to represent the degree of semantic similarity between the candidate inference step and the reference inference step. The above quantitative indicator reflects the closeness between the candidate inference step and the reference inference step in the semantic space; a higher value indicates a better semantic alignment. The alignment result is used to evaluate and filter candidate response information. When the candidate inference step and the reference inference step are highly semantically aligned, the candidate response information can be considered accurate and appropriate, and thus can be selected or adopted as the final output.

[0077] Optionally, the reference input information (such as a question in Chinese or English) and the multilingual input information (such as French, German, etc.) are linguistically different. This means that the reference input information and the input information use different language systems and vocabulary. This setting reflects a real-world scenario of cross-language processing, where the dialogue model can maintain consistency in reasoning about the same semantic content while understanding multiple languages. By comparing and aligning reasoning steps across different languages, cross-language transfer of knowledge and reasoning abilities can be facilitated while ensuring semantic integrity and logical coherence.

[0078] In this embodiment, after obtaining the candidate reasoning steps of the candidate response information, the candidate reasoning steps can be semantically aligned with the corresponding reference reasoning steps to obtain the corresponding alignment results.

[0079] Optionally, after analyzing the input information and obtaining candidate response information, this candidate response information can be decomposed into candidate reasoning steps. A cross-lingual sentence embedding model (such as LaBSE) is then used to map the candidate reasoning steps and the reference reasoning steps, transforming these two types of reasoning steps into a vector space with a unified semantic benchmark. This mapping process ensures that reasoning steps from different languages ​​can be compared at the semantic level.

[0080] Optionally, within a unified semantic representation space, the alignment result, i.e., semantic alignment, is obtained by calculating the cosine similarity between candidate inference steps and reference inference steps. Cosine similarity is a metric that measures the angular similarity between two vectors (in this scenario, the vector representations of inference steps), ranging from -1 to 1, with values ​​closer to 1 indicating higher similarity. Through this quantification method, the semantic relevance and consistency between inference steps in different languages ​​can be objectively assessed.

[0081] Optionally, based on the calculated alignment results, a comparative analysis is performed on the candidate inference steps. The overall alignment degree between each candidate inference step and the reference inference step can be evaluated. Typically, the candidate inference step sequences with the highest overall scores are selected as the multilingual responses that better conform to the source language inference path, thus ensuring the consistency of knowledge representation and inference process in cross-language scenarios. Throughout the semantic alignment process, special emphasis is placed on maintaining the performance of high-resource languages ​​(e.g., English and Chinese). This means that while improving the inference ability in low-resource languages, the original performance of the model in high-resource languages ​​is not compromised, ensuring the bidirectionality and efficiency of cross-language knowledge transfer.

[0082] In this embodiment, the above method achieves deep alignment of the cross-language reasoning process, which not only improves the accuracy and logical coherence of responses in multilingual scenarios, but also promotes knowledge sharing and capability transfer between different languages, providing strong technical support for building high-performance multilingual large models. The semantic alignment process described above is the key difference between this embodiment and related technologies such as translation-generation methods and other methods that directly rely on machine translation. Through fine-grained verification and comparison of reasoning steps, it effectively avoids the decline in reasoning ability caused by a single translation error, thereby improving the versatility and service quality of the multilingual large model in various applications.

[0083] Step S208: Based on the alignment results, determine the target response information that matches the input information from at least one candidate response information.

[0084] In the technical solution provided in step S208 of this application, the target response information can be the response that best meets the requirements and standards of the input information after semantic alignment and consistency evaluation of the reasoning process. The target response information is not only grammatically and semantically correct, but also maintains maximum consistency with the reference response in a high-resource language (such as Chinese or English) in terms of logical reasoning path. The target response information can be answered in the target language.

[0085] In this embodiment, after semantically aligning the candidate reasoning steps with the reference reasoning steps to obtain the alignment result, the target response information that matches the input information can be filtered from at least one candidate response information based on the alignment result.

[0086] Optionally, the process of filtering target response information based on alignment results is a crucial step in ensuring the output quality of the multilingual large model and the cross-lingual consistency of the inference process. Based on the acquired alignment results, an overall alignment score can be calculated for each candidate response. The overall alignment score can be calculated by averaging the alignment results of corresponding candidate inference steps, or by using a more complex weighted average that considers the importance of each inference step and its contribution to the overall answer. The goal of the above steps is to quantify the overall consistency between each candidate response and the reference input information in the inference process.

[0087] Optionally, candidate responses are ranked based on the calculated overall alignment score. Higher scores indicate that the reasoning process of a candidate response is closer to the reference reasoning steps, thus matching the input information semantically and logically. Based on the ranking, high-scoring candidate responses can be selected as target responses. In some cases, to further improve the accuracy and quality of the response, a minimum alignment score threshold can be set; only when the overall alignment score of a candidate response exceeds this threshold can it be considered a potential target response.

[0088] Optionally, based on the above sorting and filtering process, a candidate response that matches the input information in the reasoning process and has a high degree of semantic alignment can be identified as the target response. This target response not only meets the requirements of the input information in terms of syntax and semantics, but also maintains consistency with the response in the high-resource language in terms of reasoning path, ensuring the accuracy and coherence of cross-language reasoning.

[0089] In this embodiment, a filtering process based on alignment results effectively eliminates candidate responses that, while grammatically correct, have inference paths inconsistent with the source language response, thus avoiding output quality issues caused by semantic bias or logical errors. This mechanism ensures high-quality output and stable cross-language capabilities of the multilingual model when handling cross-language tasks, contributing to improved multilingual understanding and generation capabilities of the dialogue model. This allows the dialogue model to perform better in multilingual scenarios, meeting diverse user needs.

[0090] In summary, the process of filtering target response information based on alignment results ensures the accuracy and logical coherence of responses in multilingual scenarios by evaluating the semantic consistency between the reasoning steps of candidate response information and the reference reasoning steps. This is a key step in realizing cross-language ability transfer and knowledge sharing in the embodiments of this application.

[0091] Through steps S202 to S208 of this application, if it is necessary to use a dialogue model to respond to input information, the acquired input information can be input into the dialogue model, and the dialogue model can analyze the input information to obtain candidate response information. At least one candidate reasoning step corresponding to the candidate response information can be obtained; that is, the process of using the dialogue model to analyze / reason about the input information to obtain the corresponding candidate response information can be obtained. Reference reasoning steps obtained by the dialogue model analyzing reference input information that has the same semantics as the input information but a different language type can be obtained, and the candidate reasoning steps and reference reasoning steps can be semantically aligned to obtain an alignment result. Based on the alignment result, the target response information that can respond to the input information can be determined from the candidate response information. In the embodiments of this application, by semantically aligning with the reference reasoning steps, mature reasoning capabilities in high-resource languages ​​can be transferred to low-resource languages, significantly improving the universality and adaptability of the dialogue model in multilingual environments. Furthermore, by aligning the reasoning steps, consistency in the logical reasoning process of responses in different languages ​​is ensured, avoiding semantic errors or information distortion caused by language differences. This achieves the technical effect of reducing the limitations of dialogue models in processing interactive information and solves the problem of the limitations of dialogue models in processing interactive information.

[0092] The method described in this embodiment will be further described below.

[0093] As an optional implementation, the semantics of the reference input information are the same as the semantics of the input information.

[0094] In this embodiment, although the reference input information and the multilingual input information are in different languages, they express the same semantics. This means that the reference input information and the input information are logically, conceptually, or informationally equivalent, only expressed in different linguistic forms. By ensuring the semantic consistency between the reference input information and the multilingual input information, a cross-lingual reasoning standard can be constructed. This setup allows for alignment and quality control of the reasoning process based on the same semantic foundation, even if there are differences in linguistic expression.

[0095] In this embodiment, the semantics of the reference input information are identical to the semantics of the input information itself; that is, semantic alignment is a core component of the cross-language reasoning ability transfer technology solution. In natural language processing, semantic information involves multiple levels, including lexical meaning, sentence structure, contextual relationships, and implicit meanings. Ensuring that the semantics of the reference input information are identical to the semantics of the input information means that they should be consistent at these levels, even if they use different languages. For reasoning ability transfer, semantic alignment is fundamental to ensuring the logical consistency of the reasoning process across different language environments. If a dialogue model learns a specific reasoning pattern in a Chinese or English environment, then that pattern needs to be reproducible in the target language to guarantee the accuracy and reliability of cross-language reasoning.

[0096] Optionally, language differences may lead to different expressions of the same concept in different languages, or even ambiguity. In this embodiment, semantic alignment can reduce the negative impact of such differences, ensuring that the model's understood and generated answers are accurate and unambiguous in different languages.

[0097] As an optional implementation, step S204, obtaining at least one candidate reasoning step of candidate response information, includes: determining the structural information of the candidate response information, wherein the structural information of the candidate response information is used to represent the statement structure of the candidate response information; and decomposing the candidate response information according to the structural information of the candidate response information to obtain the candidate reasoning step.

[0098] In this embodiment, during the process of obtaining candidate reasoning steps for candidate response information, the structural information of the candidate response information can be determined. Based on this structural information, the candidate response information is decomposed to obtain candidate reasoning steps. The structural information of the candidate response information can represent its sentence structure. Structural information refers to the internal organization and logical relationships of the candidate response information, including the order of sentences, paragraph divisions, the use of logical connectors, and the relationship between topic sentences and detail sentences. Structural information can be used to identify and understand the composition of candidate response information (answers), distinguish each candidate reasoning step, and grasp the logical flow between them. In the reasoning alignment process of the multilingual large model, structural information acts as a bridge, allowing the multilingual large model to understand the internal logic and hierarchical structure of the candidate response information, thereby decomposing the candidate response information into a series of independent and complete reasoning steps. Each reasoning step represents a part of the logic or information in the candidate response information. This decomposition method ensures that each step can be evaluated and aligned independently, without being affected by other parts.

[0099] Optionally, structural information can be obtained through a series of natural language processing techniques, including but not limited to: syntactic analysis, which uses syntactic analysis techniques to identify the basic components and structure of the candidate response information, helping to determine which words and phrases constitute independent reasoning units; logical flow analysis, which analyzes the use of conjunctions and transitional sentences in the candidate response information, as well as the causal, conditional, and adversative logical relationships between sentences, thereby defining the boundaries of reasoning steps; topic sentence identification, which identifies the topic sentence and supporting details in longer candidate responses and decomposes them according to the hierarchical structure of the topic sentence and details; and paragraph segmentation, which can divide the candidate response information into multiple paragraphs or reasoning blocks based on clues such as line breaks, punctuation marks, or topic shifts.

[0100] It should be noted that the above-described process and method for obtaining structural information are merely illustrative examples and are not intended to impose specific limitations. Any process or method that can perform detailed analysis of the statement structure of candidate response information to decompose it into a series of candidate reasoning steps is within the protection scope of this application's embodiments.

[0101] Optionally, the process of obtaining at least one candidate reasoning step of candidate response information involves determining the structural information of the candidate response information and logically decomposing the response based on this information to obtain clear reasoning steps.

[0102] Optionally, determining the structural information is a key step in the above embodiments. The focus is primarily on understanding the internal organization of the candidate response information, including the order of statements, logical relationships, and the hierarchical structure involved in the reasoning process.

[0103] Optionally, after determining the structural information, the candidate response information can be decomposed based on the structural information to obtain candidate reasoning steps. The above decomposition process is based on structural information, ensuring that each candidate reasoning step is a semantically complete and logically independent reasoning unit.

[0104] In this embodiment, the decomposition process may include: sentence segmentation, that is, segmenting candidate response information according to the completeness and logical relationship of sentences to form a series of independent sentences; logical unit extraction, that is, the extraction of each logical unit aims to capture an independent reasoning step, which may be a conclusion, a hypothesis, an explanation, or a definition of a concept. When extracting logical units, it is necessary to ensure that each reasoning logical unit carries independent and complete logical information, providing a clear benchmark for subsequent semantic alignment and evaluation; information chain construction, that is, after completing the extraction of logical units, an information chain can be constructed, representing the information flow and logical relationship in the reasoning process. The construction of the information chain is used to understand the entire reasoning path, rather than isolated sentences or concepts; and candidate reasoning step generation, that is, based on the extraction of logical units and the construction of information chains, a series of candidate reasoning steps are generated, each of which is part of the logical reasoning process of the candidate response information.

[0105] It should be noted that the above process of decomposing candidate response information using structural information is only an example and is not subject to specific limitations.

[0106] In this embodiment, determining the structural information of candidate responses provides clear guidance for logical decomposition. The decomposition process based on structural information ensures that candidate responses are subdivided into a series of clear, logically independent candidate reasoning steps, laying a solid foundation for subsequent reasoning alignment and evaluation. This process not only helps improve the consistency of reasoning across different languages ​​in the dialogue model but also enhances its cross-language transferability and the accuracy and reliability of multilingual services. By analyzing and utilizing structural information, the inherent logical structure of responses can be more effectively identified and understood, thus maintaining the coherence of the reasoning process and the accuracy of semantics in different linguistic environments.

[0107] As an optional implementation, the candidate response information is decomposed according to its structural information to obtain candidate reasoning steps, including: decomposing the candidate response information according to its structural information to obtain multiple reasoning steps, wherein the reasoning steps include the reasoning logic of the dialogue model in the process of analyzing the input information; and identifying the key reasoning steps among the multiple reasoning steps as candidate reasoning steps, wherein the key reasoning steps have an importance to the candidate response information greater than an importance threshold.

[0108] In this embodiment, during the decomposition of candidate response information according to its structural information, the candidate response information can be analyzed to obtain multiple reasoning steps. The key reasoning steps among these steps are then identified as candidate reasoning steps. These reasoning steps can include the reasoning logic of the dialogue model during the analysis of input information. That is, a reasoning step refers to each independent logical reasoning or information processing step (a complete reasoning unit) taken by the dialogue model in the process of analyzing input information and generating candidate response information. Each reasoning step typically corresponds to a statement or information fragment in the candidate response information, containing the reasoning logic for answering the question, such as defining concepts, applying rules, explaining phenomena, and providing evidence. Reasoning steps are not limited to the part that directly answers the question; they may also include background information supporting the answer, assumptions in the reasoning process, and explicit expressions of logical relationships. Reasoning steps are a manifestation of the thought process of the dialogue model when processing input information and generating responses. By decomposing candidate response information into a series of reasoning steps, the logical coherence and semantic correctness of each reasoning step can be checked and verified, rather than just the grammar and surface meaning of the overall response. This provides a foundation for aligning subsequent reasoning processes and selecting candidate reasoning steps.

[0109] Key reasoning steps are considered more important than an importance threshold to the candidate response. In other words, key reasoning steps are those that decisively or significantly influence the understanding and generation of the candidate response. They directly answer the key points of the question or provide essential explanations and support. Without these key reasoning steps, the candidate response may lose its coherence or accuracy. The identification of key reasoning steps is based on an importance threshold, a pre-defined indicator used to measure the importance of the reasoning steps in the overall response.

[0110] Optionally, decomposing candidate reasoning steps according to the structural information of candidate responses is not only a refined processing of the internal logical structure of the model-generated responses, but also a key step in achieving high-quality cross-language ability transfer.

[0111] Optionally, decomposing candidate response information using structural information is to extract logical reasoning units, or reasoning steps, from the response in order to conduct more in-depth semantic and logical analysis.

[0112] Optionally, the dialogue model decomposes candidate response information into a series of reasoning steps based on its structural information. These decomposed reasoning steps provide the basis for subsequent fine-grained analysis and alignment. Each reasoning step is evaluated individually to ensure that it is not only grammatically correct but also logically and semantically consistent with the response in the source language.

[0113] Optionally, after acquiring multiple reasoning steps, key reasoning steps can be identified and determined. This process involves evaluating the importance of each reasoning step, typically determined by its logical position in the answer, its information content, and its contribution to the overall coherence and correctness of the answer. The importance threshold is a pre-defined numerical standard used to filter out reasoning steps that are crucial to understanding the candidate answer's information or contribute significantly to its accuracy. Only reasoning steps whose importance exceeds the importance threshold can be identified as key reasoning steps.

[0114] Optionally, the selection of key reasoning steps aims to focus on those crucial parts that carry core information and logic, ensuring the accurate transmission and understanding of these key parts in cross-language scenarios. This is because some details or auxiliary information may not be easily translated or understood accurately across different languages, but key reasoning steps are often the foundation for understanding the answer and must maintain their accuracy and consistency.

[0115] For example, the process of screening key reasoning steps involves assessing the impact of each reasoning step on the information in candidate responses, identifying those reasoning steps with an importance greater than a certain threshold as key reasoning steps. The assessment of importance can be based on various factors, including but not limited to: information contribution rate, i.e., the importance of the information contained in the reasoning step to answering the question; logical load, i.e., the role of the reasoning step in the overall logical chain and how it affects the correctness of subsequent reasoning; frequency analysis, i.e., the frequency of the reasoning step in high-resource language responses, with high-frequency reasoning steps potentially being more critical; and semantic depth, i.e., the depth and complexity of the semantic understanding of the reasoning step, with deeper reasoning steps potentially being more critical.

[0116] It should be noted that the above-described process and method for selecting key reasoning steps from the reasoning steps are merely illustrative examples and are not intended to impose specific limitations. Any method capable of selecting key reasoning steps of high importance is within the protection scope of the embodiments of this application, and will not be listed here one by one.

[0117] In this embodiment, the core of the method lies in decomposing the model-generated answer into reasoning steps and selecting key components to improve the efficiency and accuracy of cross-language ability transfer. By using structural information for decomposition, logical units in the answer can be accurately identified, while the identification of key reasoning steps ensures the priority alignment of important information and logic. This method not only improves the performance of multilingual large models in handling cross-language tasks but also ensures the consistency of the cross-language reasoning process, further enhancing the model's ability to serve the diverse needs of global users.

[0118] In summary, the process of decomposing candidate response information according to its structural information to obtain key reasoning steps is a crucial step in achieving cross-language reasoning alignment and capability transfer in the technical solution of this application. Through structured and fine-grained analysis, the model's ability to understand, generate, and align in multilingual environments is effectively improved, thereby significantly enhancing the model's versatility and output quality.

[0119] As an optional implementation, determining the structural information of the candidate response information includes: determining the information content of the candidate response information and the information content of the reference response information corresponding to the input information; determining the difference information content between the information content of the candidate response information and the information content of the reference response information; and determining the structural information of the candidate response information in response to the difference information content being less than the information content threshold.

[0120] In this embodiment, during the process of determining the structural information of candidate response information, the information content of the candidate response information and the information content of the reference response information corresponding to the input information can be determined. The difference in information content between the two can be determined. The relationship between the difference in information content and the information content threshold can be judged. If the difference in information content is less than the information content threshold, the structural information of the candidate response information can be determined. The information content of the candidate response information can be the length of the response that ensures the structural rationality of the candidate response. That is, the information content of the candidate response information can refer to the length of content required for the candidate response to express a complete and reasonable answer, and can be a quantitative indicator such as the number of words, phrases, or sentences. In this embodiment, the information content focuses more on the structural rationality of the answer, that is, whether the answer is suitable in length to fully express its core logic and information.

[0121] Optionally, the information content of the reference response can be the length of the source language response. The information content of the reference response is the amount of information contained in the source language response (usually a response in a high-resource language such as English or Chinese), and can also be expressed using metrics such as word count, sentence count, or paragraph count. The information content of the reference response provides a benchmark for the information content of candidate responses.

[0122] Optionally, the differential information content refers to the difference between the information content of the candidate response and the information content of the reference response. This difference can be an absolute difference (a direct difference in information content) or a relative difference (the ratio of the information content difference to the reference information content). Calculating the differential information content is a key step in assessing the similarity between the candidate response and the source language answer in terms of information expression. The smaller the differential information content, the closer the candidate response is to the reference response in terms of length and information density, which means that its structure is more reasonable and it is more likely to correctly and completely express the answer to the question.

[0123] Optionally, the information content threshold is a preset threshold value used to judge whether the information content of the candidate response is reasonable. If the difference in information content is less than the information content threshold, then the candidate response is considered structurally reasonable; conversely, if the difference in information content exceeds the information content threshold, the candidate response may differ significantly from the reference response in length or information density, requiring further evaluation or elimination. Setting the information content threshold is a key indicator for filtering out candidate responses that are structurally reasonable and have appropriate information content. The information content threshold can be used to exclude responses that are too short (potentially omitting important information) or too long (potentially containing redundant information), ensuring the accuracy and effectiveness of subsequent reasoning steps.

[0124] Alternatively, this implementation focuses primarily on determining the structural rationality of candidate response information through information content analysis, thereby ensuring the consistency of the cross-language reasoning process.

[0125] Optionally, the length and content density, or information content, of each candidate response can be evaluated. Information content can be measured by word count, sentence count, or sentence count, but it represents a quantitative indicator of how much content a candidate response needs to express a complete and reasonable answer. This evaluation process helps to understand the complexity and completeness of candidate responses, providing a basis for logical decomposition in subsequent steps. Evaluating the information content of candidate responses ensures they are sufficiently complete to contain the key information needed to answer the question, while avoiding excessive length and interference from irrelevant information. Information content evaluation is a preliminary step in screening responses of reasonable length, helping to eliminate candidate responses that are too incomplete (potentially omitting key points) or too incomplete (potentially containing redundant or irrelevant information).

[0126] Optionally, reference responses refer to responses using a high-resource language (such as English or Chinese) to answer the same question, providing a benchmark for the information content of candidate responses. The information content of reference responses can be evaluated to determine a reasonable standard for response length and information density, serving as the basis for subsequent comparison and selection. The role of reference information content is to serve as a benchmark for measuring the reasonableness of the information content of candidate responses. By comparing the information content of candidate responses and reference responses, it can be determined whether candidate responses are consistent with the reference response in terms of length and information density, thereby ensuring the logical coherence and completeness of responses in a cross-language context.

[0127] Optionally, the differential information content is the difference between the information content of the candidate response and the information content of the reference response. Calculating this differential information content is to assess how closely the candidate response is similar to the reference response in terms of information content, that is, to determine whether it maintains consistency with the response in the high-resource language in terms of length and information density. The magnitude of the differential information content directly reflects the structural rationality of the candidate response in reconstructing the target language response.

[0128] Optionally, after determining the amount of difference information, the amount of difference information can be compared with a preset information content threshold. If the amount of difference information is less than the information content threshold, it means that the candidate response information is close enough to the reference response information in terms of information content, and its structural information can be determined for subsequent reasoning steps. Conversely, if the amount of difference information is too large, it indicates that the candidate response information may be too long or too short, and there are unreasonable aspects in its structure, requiring further evaluation or elimination.

[0129] In this embodiment, the above method effectively filters out candidate response information that meets structural requirements, ensuring that the information content of the candidate response information is consistent with that of the high-resource language response. The implementation of the above process considers not only the length of the candidate response information but also the rationality of the information content, thereby avoiding further processing of extremely long responses and improving the accuracy and efficiency of subsequent reasoning alignment. By comparing the difference in information content with an information content threshold, a preliminary assessment of the structural rationality of the candidate response information is achieved, providing a reliable screening standard for subsequent reasoning alignment and candidate reasoning step generation.

[0130] In summary, the above implementation methods enhance the ability of multilingual large models to output reasoning information in cross-lingual scenarios by quantifying information content and comparing differences, ensuring the consistency and accuracy of logical reasoning and information expression in responses generated in different language environments. This not only helps improve the model's performance in low-resource languages ​​but also ensures that the model's reasoning performance in high-resource languages ​​is not compromised, achieving effective transfer and sharing of cross-lingual reasoning capabilities.

[0131] As an optional implementation, the language type of the input information is the target language type, and the language type of the reference input information is the source language type. The dialogue model is trained based on information samples in the target language type and information samples in the source language type, and the sample size of the information samples in the target language type is lower than the sample size of the information samples in the source language type.

[0132] In this embodiment, the language type of the input information is the target language type, which can be a low-resource language, such as a multilingual query other than Chinese / English. The target language type refers to the language type used to generate candidate response information in a large multilingual model. In this embodiment, the target language type refers to a low-resource language, that is, a language type with less data and fewer resources compared to high-resource languages ​​(such as English and Chinese). For example, Japanese, Korean, Spanish, and Arabic have fewer instruction data in existing training datasets, so the performance of the dialogue model on these languages ​​may be inferior to that of high-resource languages. Aligning the inference process under the target language type is the core objective of this technical solution. By improving its consistency with the inference process of high-resource languages, the inference capabilities of the large model can be effectively transferred and shared, thereby maintaining and improving performance even in a low-resource language environment.

[0133] Optionally, the language type of the reference input information is the source language type, which can be a high-resource language, such as a Chinese / English query. The source language type refers to the language type that provides rich instruction data during the training of a multilingual large-scale model; it is typically a high-resource language, such as English or Chinese. The reasoning process and responses under the source language type are considered as benchmarks for understanding and generation capabilities, and the model's performance on these languages ​​tends to be more stable and reliable. The reference reasoning steps under the source language type serve as alignment benchmarks for evaluating and adjusting the reasoning process under the target language type, ensuring accurate transfer of knowledge and capabilities between different languages. Through training on datasets under the source language type, the reasoning patterns and strategies formed by the model when processing source language instructions can be used as a reference for optimizing the reasoning process under the target language type.

[0134] Optionally, the semantic space is a high-dimensional vector space where each point represents the semantic features of a sentence or phrase. In this semantic space, the smaller the distance between two sentences, the more similar their semantics; conversely, the larger the distance, the greater the semantic difference. The construction of the semantic space typically relies on pre-trained cross-lingual sentence embedding models, such as LaBSE, which can map sentences from different languages ​​to the same semantic space, enabling cross-lingual semantic representation and comparison. Mapping candidate inference steps in the target language type and reference inference steps in the source language type to the same semantic space is a key step in achieving fine-grained alignment of the inference process. In the semantic space, the semantic alignment between two inference steps can be quantitatively evaluated by calculating metrics such as cosine similarity, thereby selecting the target language answer that is most semantically consistent with the source language answer, ensuring the logical consistency and semantic accuracy of the cross-lingual inference process.

[0135] As an optional implementation, step S206 involves semantically aligning the candidate reasoning steps with the corresponding reference reasoning steps to obtain an alignment result, including: mapping the candidate reasoning steps under the target language type and the reference reasoning steps under the source language type to the semantic space respectively; and semantically aligning the mapped candidate reasoning steps and the mapped reference reasoning steps to obtain an alignment result.

[0136] In this embodiment, during the semantic alignment of candidate inference steps with their corresponding reference inference steps, the candidate inference steps in the target language type and the reference inference steps in the source language type can be mapped to the same semantic space. The mapped candidate inference steps and the mapped reference inference steps can then be semantically aligned to obtain the alignment result.

[0137] Alternatively, this implementation is a technical solution achieved through fine-grained inference step alignment in the context of cross-language ability transfer, in order to improve the understanding and generation capabilities of multilingual large models in the target language type (usually a low-resource language).

[0138] Optionally, candidate inference steps in the target language type and reference inference steps in the source language type are mapped to a unified semantic space. This process relies on a pre-trained cross-lingual sentence embedding model, such as LaBSE, capable of processing texts from different languages ​​and converting both inference steps into high-dimensional vectors. These high-dimensional vectors capture the semantic features of the sentences in the inference steps, rather than just their surface textual form. The purpose of mapping to the semantic space is to enable inference steps from different languages ​​to be compared and analyzed at the semantic level, without being limited by specific language vocabulary and grammatical structures.

[0139] Optionally, after semantic space mapping is completed, the mapped candidate inference steps can be semantically aligned with the mapped reference inference steps. This can be achieved by calculating the similarity between the vectors of the two candidate inference steps in the semantic space, for example, by calculating the cosine similarity between the two vectors.

[0140] Optionally, alignment results can be obtained by calculating the cosine similarity between each pair of candidate inference steps and the reference inference step. For each candidate inference step, a suitable matching reference inference step under the source language type can be found, that is, the inference step with high similarity between the two in the semantic space. The above similarity not only reveals the semantic similarity between inference steps, but also provides a quantitative basis for subsequent decision-making.

[0141] Optionally, after obtaining the alignment results of the inference step pairs (candidate inference steps and their corresponding reference inference steps), decisions can be made based on the alignment results. For example, if the average cosine similarity between the candidate inference step and the reference inference step is higher than a certain preset threshold, or if the similarity on several key steps exceeds a specific threshold, then the candidate inference step will be considered to be semantically well aligned with the reference inference step. The above alignment results can be used to select suitable multilingual answers or to guide the model to optimize the inference process for specific language types, thereby improving its understanding and generation capabilities.

[0142] In this embodiment, by meticulously comparing and aligning the reasoning processes between different languages ​​at the semantic level, comprehension and generation biases caused by language differences can be effectively identified and corrected, thereby maintaining high model performance even in low-resource language environments. This not only enhances the model's cross-lingual versatility but also promotes the effective transfer and sharing of knowledge and reasoning abilities across different languages, making it a key method for achieving high-level performance of large multilingual models in multilingual environments.

[0143] As an optional implementation, the candidate inference steps under the target language type and the reference inference steps under the source language type are mapped to the semantic space, respectively. This includes: inputting the candidate inference steps under the target language type and the reference inference steps under the source language type into an information conversion model, wherein the information conversion model is obtained by training an encoder; and using the information conversion model, converting the candidate inference steps under the target language type and the reference inference steps under the source language type into vectors in the semantic space, wherein the dimension of the vectors in the semantic space is greater than a dimension threshold.

[0144] In this embodiment, during the process of mapping candidate inference steps under the target language type and reference inference steps under the source language type to the semantic space, the candidate inference steps and reference inference steps can be input into an information conversion model. The information conversion model is then used to convert the candidate inference steps and reference inference steps into vectors in the semantic space. The information conversion model can be trained on an encoder and can be a cross-language sentence embedding model. The encoder can be BERT. The vector dimension in the semantic space is greater than a dimensionality threshold; that is, the semantic space is a high-dimensional semantic space.

[0145] Optionally, this embodiment is a key technology for achieving cross-language reasoning process alignment and capability transfer. By mapping candidate reasoning steps and reference reasoning steps to a high-dimensional semantic space, it realizes the unification and comparison of semantic representations between different languages.

[0146] Optionally, the information conversion model is a key component for converting text information from different languages ​​into a unified semantic vector representation. In this technical solution, the information conversion model is typically trained on an encoder, which can be based on the BERT architecture. BERT is a powerful pre-trained model capable of simultaneously considering both forward and backward information from the context to generate high-quality text vector representations. Further training on a large amount of multilingual text data can yield a model specifically designed for cross-lingual semantic mapping, known as a cross-lingual sentence embedding model (such as LaBSE).

[0147] Optionally, candidate inference steps in the target language type and reference inference steps in the source language type are input into the information conversion model, respectively. The information conversion model, based on the trained weights and algorithm, converts the input inference steps into high-dimensional vectors. These high-dimensional vectors accurately represent the semantic features of the original sentence in the semantic space. In this way, inference steps from different languages ​​are uniformly represented, providing a foundation for subsequent semantic alignment.

[0148] Optionally, through an information transformation model, candidate inference steps and reference inference steps are converted into vector representations in this space. Setting a dimensionality threshold ensures the richness and granularity of the semantic representation, enabling more accurate capture of sentence semantic details while handling semantic differences between languages. Mapping to a high-dimensional semantic space allows inference steps that would otherwise exist in different linguistic contexts to be compared within a unified semantic framework. This not only eliminates differences in linguistic form but also ensures semantic alignment; even if different languages ​​have different expressions, as long as the core semantics are similar, close vector representations can be found in the high-dimensional semantic space.

[0149] Optionally, in a high-dimensional semantic space, each reasoning step can be represented as a vector containing its semantic features. By calculating the similarity (e.g., cosine similarity) between different vectors, the semantic proximity between candidate and reference reasoning steps can be quantitatively assessed. This process is crucial for identifying and correcting semantic biases in cross-linguistic reasoning. The advantage of high-dimensional vector representation is its ability to capture subtle semantic differences within sentences, and by calculating distance or similarity, it can effectively find semantically aligned reasoning steps across different languages. This vectorized representation method is the cornerstone of cross-linguistic competence transfer and reasoning process alignment, ensuring accurate understanding and generation of semantically equivalent reasoning steps even in low-resource language environments.

[0150] In the embodiments of this application, the above implementation maps the reasoning steps of different languages ​​to a high-dimensional semantic space and uses an information transformation model (such as a BERT-based cross-lingual sentence embedding model) to capture and unify semantic representations, thereby achieving consistency and alignment in the cross-lingual reasoning process. This method not only overcomes the problem of insufficient low-resource language data but also promotes the sharing and transfer of reasoning capabilities between different languages, significantly improving the performance and versatility of multilingual large models in cross-lingual scenarios.

[0151] As an optional implementation, semantic alignment is performed on the mapped candidate inference steps and the mapped reference inference steps to obtain the alignment result, including: determining the semantic similarity between the vectors of the candidate inference steps in the target language type and the vectors of the reference inference steps in the source language type in the semantic space, wherein the vector dimension in the semantic space is greater than the dimension threshold, and the semantic similarity is used to represent the alignment result.

[0152] In this embodiment, during the process of semantically aligning the mapped candidate inference steps and the mapped reference inference steps to obtain the alignment result, the semantic similarity between the vectors of the candidate inference steps and the reference inference steps in the semantic space can be determined. The semantic similarity can be used to represent the alignment result. The semantic similarity can be cosine similarity, which measures the semantic alignment between the candidate inference steps and the reference inference steps. The dimension of the vectors in the semantic space is greater than a dimension threshold; that is, the semantic space is a high-dimensional vector space.

[0153] Optionally, this embodiment is a core component of cross-language reasoning alignment technology, the goal of which is to ensure semantic consistency between the reasoning process in the target language (low-resource language) and the reasoning process in the source language (high-resource language).

[0154] Optionally, in the cross-language reasoning alignment process, candidate reasoning steps in the target language type and reference reasoning steps in the source language type are mapped to a semantic space. By using a cross-language sentence embedding model (such as LaBSE), the reasoning steps are converted into vectors in the semantic space, achieving uniformity and comparability of semantic representations across different languages. This conversion process is based on the model's ability to understand large amounts of text, capturing the core semantic features of the reasoning steps, rather than just the surface textual or grammatical structures.

[0155] Optionally, calculating semantic similarity is a crucial step in evaluating the alignment of inference steps. Semantic similarity can be calculated in various ways; in this embodiment, cosine similarity is an effective method. Cosine similarity measures the semantic closeness of two vectors by calculating the cosine of the angle between them. Its value ranges from -1 to 1, with values ​​close to 1 indicating that the two vectors are very similar, while values ​​close to 0 or -1 indicate that they are semantically significantly different.

[0156] For example, in cross-language reasoning alignment, the specific process of calculating semantic similarity is as follows: For each candidate reasoning step in the target language and its corresponding reference reasoning step in the source language, they are first converted into vector representations in the semantic space. The cosine similarity between the target language reasoning step vector and the source language reference reasoning step vector is calculated. This process may include multiplying the two vectors by their respective moduli and then dividing by the product of their magnitudes. Finally, the resulting cosine similarity value reflects the semantic alignment between the two reasoning steps; a larger cosine similarity value indicates a higher degree of alignment.

[0157] Optionally, semantic similarity is directly used to represent the alignment result, providing a quantitative indicator of semantic consistency between candidate inference steps and reference inference steps. By calculating the semantic similarity between candidate inference steps and their corresponding reference inference steps, an alignment matrix can be constructed, where each element represents the degree of alignment between a pair of inference steps. Based on the alignment matrix, candidate responses with high overall inference path scores can be selected, i.e., multilingual responses that are semantically closest to the source language response (reference inference step) in terms of candidate inference steps.

[0158] In this embodiment of the application, in the above-described cross-language reasoning alignment implementation, by mapping candidate reasoning steps and reference reasoning steps to a semantic space and calculating the cosine similarity between the two reasoning steps, the semantic alignment of reasoning steps in different languages ​​can be accurately evaluated and quantified. This mechanism ensures that the multilingual large model maintains consistent reasoning logic and language understanding capabilities in low-resource language environments as it does in high-resource language environments, thereby improving the versatility of the dialogue model and its performance in diverse language scenarios. Fine-grained semantic alignment not only enhances the consistency of cross-language reasoning in the model but also promotes the effective transfer of knowledge and abilities between different languages, making it a key technology for achieving high-quality multilingual services.

[0159] As an optional implementation, at least one candidate inference step is used to construct the inference path for the dialogue model to analyze the input information. Based on the alignment result, the target response information that matches the input information is determined from at least one candidate response information. This includes: determining the evaluation result of the inference path based on the semantic similarity corresponding to the candidate inference step, wherein the semantic similarity is used to represent the alignment result; and determining the target response information from at least one candidate response information based on the evaluation result.

[0160] In this embodiment, during the process of filtering target response information from candidate response information based on the alignment result, the evaluation result of the reasoning path can be determined based on the semantic similarity corresponding to the candidate reasoning steps. The target response information can then be determined from the candidate response information based on the evaluation result. The candidate reasoning steps can be used to construct the reasoning path (response path) for the dialogue model to analyze the input information. The evaluation result of the reasoning path can be the reasoning path score, i.e., the overall reasoning path score. Semantic similarity is used to represent the alignment result; that is, the alignment result can be the semantic similarity between the semantics of the candidate reasoning step and the semantics of the reference reasoning step.

[0161] Optionally, this implementation aims to optimize the selection of responses in multilingual large models in cross-language scenarios by quantifying the semantic similarity of inference paths, ensuring that the responses generated by the model are not only grammatically correct in different language environments, but also logically coherent and semantically accurate.

[0162] Optionally, candidate reasoning steps constituting the reasoning path can be extracted from the multiple candidate response information generated by the multilingual large model. These candidate reasoning steps represent the logical reasoning process undertaken by the large model when processing input information, with each step corresponding to a specific aspect or level of logical reasoning within the input information. For example, if the input information is a question, the candidate reasoning steps might include understanding the question, retrieving relevant knowledge, unfolding the logical reasoning process, and generating the final answer.

[0163] Optionally, based on the candidate reasoning steps decomposed above and their vector representations in the semantic space with the reference reasoning steps of the source language (high-resource language), the semantic similarity between each pair of reasoning steps is calculated. Based on the semantic similarity corresponding to each candidate reasoning step, the quality of the entire reasoning path is comprehensively evaluated, generating a reasoning path score. The reasoning path score reflects the degree of consistency between the entire answer and the source language answer in terms of logical reasoning and semantic expression, and is an important indicator for measuring the effectiveness of cross-language reasoning alignment.

[0164] Optionally, based on the evaluation results of the reasoning path (i.e., the reasoning path score), the target response information that matches the input information is selected from multiple candidate response information. This process can be understood as an optimization problem, with the goal of minimizing the semantic bias of the cross-lingual reasoning process and maximizing the overall score of the reasoning path. The candidate response information with the highest reasoning path score is selected as the target response information, i.e., the multilingual response that is semantically closest to the source language response.

[0165] Optionally, the aforementioned screening mechanism, through fine-grained inference step alignment and scoring, can effectively identify and control semantic biases in cross-language inference processes. This ensures that the responses generated by the model in different language environments not only conform to the norms of the target language in terms of grammar and vocabulary, but also maintain a high degree of consistency with the source language in terms of inference logic and semantic expression. In this way, not only is the performance of the multilingual large model improved on low-resource languages, but its inference ability on high-resource languages ​​is also ensured to remain unimpaired, achieving effective transfer and sharing of knowledge and capabilities in cross-language scenarios.

[0166] In the embodiments of this application, the above method can be a multilingual large model response selection mechanism based on the consistency of the reasoning process. By quantitatively evaluating the semantic alignment of each reasoning path, it ensures that the dialogue model can generate high-quality and highly consistent answers when dealing with multilingual scenarios, thereby significantly enhancing the model's cross-language versatility and user service experience.

[0167] As an optional implementation, based on the evaluation results, determining the target response information from at least one candidate response information includes: selecting candidate response information whose evaluation results are higher than the evaluation result threshold from at least one candidate response information; and determining the selected candidate response information as the target response information.

[0168] In this embodiment, during the process of determining the target response information from the candidate response information based on the evaluation results, candidate response information with evaluation results higher than the evaluation result threshold can be selected. Thus, the selected candidate response information with evaluation results higher than the evaluation result threshold is determined as the target response information. That is, the target language answer with a high overall reasoning path score can be selected from the candidate response information as the final output, realizing knowledge transfer from reference input information to input information.

[0169] Optionally, this implementation method filters and determines the target response information by setting an evaluation result threshold, ensuring that the response output by the dialogue model is not only semantically consistent with the source language, but also meets a certain standard in terms of quality.

[0170] Optionally, the evaluation result threshold is a key parameter for filtering candidate response information, used to determine whether the score of the candidate reasoning path is high enough to meet the criteria for a high-quality response. The evaluation result threshold is set based on the performance of the multilingual large model during training and validation, and can be an empirical value aimed at balancing the accuracy and speed of multilingual understanding and generation. The evaluation result threshold can be set relatively high to ensure that the output multilingual response is highly consistent with the source language response in terms of reasoning process and semantic expression, thereby improving the quality of the response; or it can be set relatively moderately to increase the flexibility and efficiency of model processing while ensuring a certain level of quality.

[0171] Optionally, after calculating the semantic similarity between the candidate reasoning step and the reference reasoning step, and obtaining the evaluation result (reasoning path score) for each reasoning path based on this, the above reasoning path score can be compared with a preset evaluation result threshold. Candidate response information with a score higher than the threshold is considered as a high-quality candidate that is well aligned with the source language response in terms of reasoning process and semantics, and the above candidate response information will be further considered as the target response information for the final output.

[0172] Optionally, after filtering out candidate responses with reasoning path scores higher than the evaluation result threshold, the target response is further determined from these high-quality candidates. For example, the candidate response with the highest reasoning path score can be selected as the response in the target language, i.e., the final multilingual response. This process ensures that the output response not only conforms to the norms of the target language in terms of grammar and language expression, but also maintains a high degree of consistency with the reference response in the source language in terms of logical reasoning and semantic expression, thereby achieving high-quality cross-language knowledge transfer.

[0173] In the embodiments of this application, the answers generated by the multilingual large model in the target language environment are not only grammatically and lexically correct, but also highly consistent with the answers in the source language in terms of logical reasoning and semantic understanding. This consistency is key to cross-language ability transfer, meaning that the large model can effectively transfer reasoning abilities and knowledge from the source language to the target language, thereby maintaining a high level of service quality in different language environments and meeting the multilingual needs of global users.

[0174] In summary, setting evaluation result thresholds and using these thresholds to filter and determine target response information is a crucial mechanism for ensuring the alignment effect and response quality of cross-language reasoning in multilingual large-scale models. The aforementioned strategies, by controlling the overall quality of the reasoning path, effectively promote the transfer of knowledge and skills across different languages, providing users in multilingual scenarios with more accurate, natural, and consistent responses, and improving the service quality of multilingual large-scale models.

[0175] As an optional implementation, step S102, obtaining input information, includes: obtaining initial input information in the source language type; converting the initial input information from the source language type to the target language type to obtain input information, wherein the dialogue model is trained based at least on information samples in the target language type and information samples in the source language type, and the sample size of information samples in the target language type is lower than the sample size of information samples in the source language type.

[0176] In this embodiment, during the process of acquiring input information, initial input information in the source language type can be obtained. This initial input information can be converted from the source language type to the target language type to obtain the final input information. The dialogue model can be trained based on information samples in both the target language type and the source language type. The sample size of the information samples in the target language type is smaller than that in the source language type. The initial input information in the source language type can be a question in the source language (Chinese / English query).

[0177] Optionally, this implementation describes how to transform and acquire input information during cross-language ability transfer to facilitate the analysis and generation of a multilingual large model.

[0178] Optionally, the initial input information in the source language type can be a query from the user or system, presented in a high-resource language (source language), such as a question in Chinese or English (Chinese / English query). The initial input information in the source language type is the foundation for the dialogue model's understanding and generation, and can contain the content the user wants to know or the service requested. Since the goal of this application embodiment is to enhance the performance of the multilingual large model in a low-resource language environment, the obtained source language query can be automatically translated into the target language. The above process uses machine translation technology or cross-language mapping mechanisms to transform the Chinese or English question into a question in another language (such as Japanese, Arabic, etc.), obtaining the input information in the target language type. The choice of target language depends on user needs or specific application scenarios.

[0179] Optionally, the dialogue model is trained on multilingual text data, including information samples in both the source and target languages. However, due to the scarcity of data in low-resource languages, the number of information samples in the target language is typically much smaller than that in the source language. This results in the model performing worse on low-resource languages ​​than on high-resource languages. Therefore, by transforming the source language problem into a target language problem, the rich training data in the source language can be used to optimize and improve the model performance in the target language, achieving cross-lingual capability transfer and alignment.

[0180] Optionally, the initial input information in the source language is converted into input information in the target language, aiming to enable the multilingual large model to more effectively understand and generate responses in the target language environment. Although the training sample size for the target language is smaller, by leveraging the rich dataset and training experience of the source language, the multilingual large model can generate responses in the target language based on the reasoning path and language understanding ability of the source language, thereby providing high-quality services even in low-resource language environments.

[0181] In this embodiment, the above implementation emphasizes that in cross-language scenarios, the initial input information in the source language is translated and converted into target language input information that the multilingual large model can understand and process. This allows the multilingual large model to leverage its accumulated knowledge and experience in high-resource languages ​​to improve its performance in low-resource language scenarios. This conversion process is a prerequisite for achieving cross-language capability transfer and inference path alignment, ensuring that the multilingual large model can provide accurate, natural, and consistent responses in different language environments.

[0182] As an optional implementation, the method further includes: determining the initial input information under the source language type as reference input information; inputting the reference input information into the dialogue model for analysis to obtain reference response information under the source language type.

[0183] In this embodiment, the initial input information in the source language type can be determined as the reference input information. This reference input information can be input into the dialogue model, and the dialogue model can analyze the reference input information to obtain the reference response information in the source language type. The reference response information can be a response in the source language.

[0184] Alternatively, this implementation removes the framework for cross-language ability transfer and further enhances the accuracy and consistency of the model in generating responses in the target language by explicitly distinguishing and utilizing input and response information from the source language.

[0185] Optionally, before starting cross-language ability transfer, the input information in the source language can be clearly defined. The source language can be a language with abundant training data, such as Chinese or English. In the above steps, the initial input information obtained under the source language type (i.e., the initial Chinese or English question posed by the user) is determined as the reference input information, which is an important benchmark for the entire cross-language reasoning alignment process.

[0186] Optionally, the reference input information can be fed into a pre-trained dialogue model. This model primarily uses information samples from the source language type as training data, while also incorporating a smaller number of information samples from the target language type. Based on its learning in the source language, the dialogue model can analyze the reference input information and generate one or more reference responses in the source language type. These reference responses are high-quality replies generated by the model in a high-resource language environment based on the source language input. They possess high standards in terms of syntax, vocabulary, logical reasoning, and semantic expression, and can serve as a quality reference for the target language response.

[0187] Optionally, after obtaining the reference response information in the source language, the next step is to compare and align it with the candidate response information generated in the target language. This process involves the aforementioned steps of decomposing reasoning steps, transforming semantic representations, and scoring reasoning paths. The aim is to ensure that the target language response maintains consistency with the reference response in logical reasoning and semantic expression, thereby achieving high-quality cross-linguistic knowledge transfer. Based on the alignment results, the candidate response information is evaluated, and the target language response that closely approximates the reference response information is selected as the final output. This selection process considers not only grammatical and lexical correctness but also emphasizes the consistency of logical reasoning and semantic expression, ensuring the quality and naturalness of the output response.

[0188] In this embodiment, the above steps effectively utilize high-quality source language data and the model's strong understanding capabilities in the source language, providing strong support for generating high-quality responses in low-resource language environments. This not only solves the problem of poor model performance due to limited target language data but also promotes knowledge sharing and reasoning ability transfer between different languages, making it one of the key strategies for enhancing the cross-linguistic versatility of multilingual large models.

[0189] As an optional implementation, the method further includes: determining the structural information of the reference response information, wherein the structural information of the reference response information is used to represent the statement structure of the reference response information; and decomposing the reference response information according to the structural information of the reference response information to obtain the reference reasoning steps.

[0190] In this embodiment, the structural information of the reference response information can be determined. This structural information can represent the statement structure of the reference response information. Based on this structural information, the reference response information can be analyzed to obtain the reference reasoning steps.

[0191] Optionally, this implementation is an important component of achieving cross-language reasoning alignment and capability transfer. By analyzing the sentence structure of the source language (high-resource language) response, the reasoning process is further refined, providing a structured reference for aligning subsequent reasoning steps.

[0192] Optionally, the reference response information is a high-quality answer generated by the dialogue model in a high-resource language environment. Structural information, mainly including the logical framework of the answer, the relationships between sentences, and the organization of key information, is the foundation for understanding the logical flow and semantic organization of the answer. The process of determining structural information can employ natural language processing techniques such as syntactic analysis, semantic role labeling, and text segmentation to decompose the reference response information into a series of coherent sentences or reasoning units. This structural information not only reflects the grammatical and lexical features of the answer but also embodies its inherent logical reasoning and semantic organization, serving as the basis for alignment and transfer in subsequent steps.

[0193] Optionally, based on the determined structural information, the reference response information is meticulously decomposed and transformed into a series of independent, semantically complete reasoning steps. This decomposition process aims to make the logical reasoning process in complex responses explicit, enabling each reasoning step to be understood and evaluated independently. For example, the explanation of a scientific concept may include several parts such as definition, principle, and application examples. Through structural information analysis, these parts can be extracted separately as independent reasoning steps.

[0194] Optionally, the decomposed reference reasoning steps are the smallest units of logical reasoning and semantic expression in the reference response information. Each reasoning step should contain complete and explicit information, capable of being semantically represented and aligned independently. Obtaining the reference reasoning steps provides a clear comparative standard for subsequent cross-language reasoning step alignment, enabling the system to identify and select candidate responses that are consistent with the source language response in the reasoning process, within the response of the target language (low-resource language).

[0195] In this embodiment, the high-quality responses in the source language are not only transformed into candidate responses in the target language using the above method, but their inherent reasoning processes are also finely structured and made explicit. These structured reference reasoning steps allow cross-language reasoning alignment to extend beyond the lexical and grammatical levels to the levels of logical reasoning and semantic expression, significantly improving the versatility and knowledge transfer efficiency of the multilingual model in cross-language scenarios. Ultimately, based on the alignment between the reference reasoning steps and the reasoning steps of the target language candidate responses, appropriate multilingual responses can be selected, achieving high-quality cross-language services.

[0196] As an optional implementation, the method further includes: adjusting the parameters of the dialogue model in response to a reference adjustment operation; inputting input information into the dialogue model for analysis to obtain at least one candidate response, including: inputting the input information into the adjusted dialogue model to obtain a number of candidate response messages corresponding to the adjusted parameters.

[0197] In this embodiment, if a parameter adjustment operation is detected, the parameters of the dialogue model can be adjusted. The adjusted dialogue model is then used to analyze the input information to obtain a number of candidate responses corresponding to the adjusted parameters. The parameters of the dialogue model can be hyperparameters, which can be used to adjust the answer space.

[0198] Alternatively, this implementation focuses on optimizing the performance of the dialogue model under specific input information by dynamically adjusting its parameters, especially in the context of cross-language ability transfer, aiming to improve the quality and diversity of multilingual candidate response information generated by the model.

[0199] Optionally, when an operation signal requiring adjustment of model parameters is detected, in order to cope with the complexity or specificity of particular input information, or to better adapt to the characteristics of the target language, the reference adjustment operation can be based on previous model generation results or user feedback to identify potential deficiencies or biases in the model in certain scenarios, thereby triggering parameter adjustments in order to achieve the desired effect in subsequent generation.

[0200] Optionally, in response to the adjustment operation, the parameters of the dialogue model will be fine-tuned. These parameters may include hyperparameters, such as learning rate, batch size, and temperature parameters, which directly affect the model's learning and generation process. Furthermore, adjustments to the model's internal weights or architecture may also be involved, especially when optimizing for specific question types or language characteristics. By adjusting the model parameters, the space in which it generates answers can be altered, increasing the diversity and quality of candidate response information and improving the model's generalization ability in cross-linguistic scenarios.

[0201] Optionally, after parameter adjustments, the input information can be re-analyzed and generated using the adjusted dialogue model. The dialogue model can generate a number of candidate responses corresponding to the adjusted parameters. The phrase "a number corresponding to the adjusted parameters" means that, based on the new parameter configuration, the dialogue model can generate more or fewer candidate responses to reflect changes in the answer space after parameter adjustments. For example, increasing the temperature parameter might increase the diversity of candidate responses, while adjusting the generation length limit might affect the number and detail of candidate responses.

[0202] Optionally, the multiple candidate responses generated based on the adjusted model will be further analyzed and filtered. Semantic similarity is calculated by aligning the reasoning steps with the source language reference responses, evaluating the coherence and accuracy of the reasoning paths, and thus selecting a suitable multilingual response as the final output. This selection process fully leverages the improved response diversity and quality resulting from parameter adjustments, ensuring that the dialogue model's performance in cross-language scenarios more closely matches the reference standard of the high-resource language.

[0203] In this embodiment, the method described above dynamically adjusts the dialogue model parameters and, combined with the model generation, analysis, and selection process, can flexibly optimize the model's performance when encountering specific input information or a particular language environment. This method not only helps improve the response quality of a large multilingual model in low-resource language environments but also ensures that the performance of a large multilingual model is not compromised in high-resource languages, making it an effective means of achieving high-quality cross-language capability transfer.

[0204] This application also provides another method for processing interactive information. Figure 3 This is a flowchart of another method for processing interactive information according to an embodiment of this application, such as... Figure 3 As shown, the method may include the following steps:

[0205] Step S302: Obtain query information in the target language type, wherein the query information is used to represent the information to be answered in the dialogue scenario.

[0206] In this embodiment, in a cross-language scenario, the query information in the target language refers to the question or request raised by the user or system in a low-resource language (target language). This is the core content that needs to be understood and the answer generated during the multilingual large-scale model processing. For example, if the source language is English or Chinese, the target language might be Japanese, Korean, or Arabic, etc., and the system can process and understand questions in the aforementioned low-resource languages.

[0207] Optionally, the target language query information constitutes the set of information to be answered in the dialogue scenario; that is, the dialogue model needs to provide corresponding answers or solutions based on the above information. This information can include inquiries about a scientific concept, seeking solutions, requesting explanations of complex phenomena, etc., covering various scenarios that a multilingual large-scale model may encounter.

[0208] Optionally, by acquiring query information in the target language type, the cross-language ability transfer method is ensured to effectively address the diverse needs of multilingual users, providing a direct input source for subsequent answer generation and inference alignment. The accuracy and efficiency of the above steps directly affect the service quality of the multilingual large model in different language environments and are key to achieving high-quality cross-language interaction.

[0209] Step S304: Input the query information into the dialogue model for analysis to obtain at least one candidate response information. The dialogue model is trained based on information samples in the target language type and information samples in the source language type. The sample size of the information samples in the target language type is lower than the sample size of the information samples in the source language type.

[0210] This embodiment illustrates how to use a dialogue model to generate candidate response information in the target language, while taking into account the difference in data volume between the target language and the source language.

[0211] Optionally, after obtaining the query information in the target language, this query information can be input into the dialogue model for analysis. The dialogue model here is a large-scale language model trained in multiple languages. It is trained not only on rich information samples from the source languages ​​(such as English and Chinese) but also considers information samples from the target language (a low-resource language), although the sample size for the target language is usually much smaller.

[0212] Optionally, the dialogue model can perform deep analysis based on the input query information, leveraging its language understanding and generation capabilities to generate at least one candidate response. The candidate response information represents the model's understanding and attempt to answer the question, reflecting its grammatical, lexical, and semantic capabilities in understanding and generating the target language.

[0213] Optionally, due to the limited information samples (training data) in the target language, the dialogue model may not perform as well in the target language as it does in the source language. However, by combining training data from both the source and target languages, the model can learn the commonalities and differences between different languages, leveraging the rich knowledge and experience of the source language to compensate for the performance degradation caused by insufficient target language data. This training strategy helps the model, when processing the target language, to draw upon the reasoning paths and knowledge of the source language, generating higher-quality candidate response information.

[0214] Optionally, when generating candidate response information, the dialogue model can employ various strategies, such as adjusting the temperature parameter or increasing the number of candidates. These strategies aim to increase the diversity of the generated responses and attempt to cover a wider range of possibilities, in order to find the most suitable answer in subsequent inference steps for alignment and filtering. By generating multiple candidate response information, a more comprehensive comparison and evaluation can be performed, and the candidate response that meets the semantic alignment requirements can be selected as the final output.

[0215] In the embodiments of this application, the above method is a key step in generating candidate response information based on the query information in the target language. It utilizes the training results of the dialogue model in a multilingual environment. In particular, when dealing with the imbalance between the amount of source language and target language data, the generation strategy and subsequent filtering steps ensure that the dialogue model can provide high-quality multilingual services.

[0216] Step S306: Obtain at least one candidate inference step of candidate response information, wherein the candidate inference step is used to represent the process by which the dialogue model analyzes the input information.

[0217] In this embodiment, the candidate reasoning step of obtaining candidate response information is a key step in the cross-language ability transfer method for deeply analyzing candidate response information. Its goal is to transform the candidate response information generated by the model into a series of candidate reasoning steps to refine and make explicit the model's decision-making process.

[0218] Optionally, the candidate response information obtained from the dialogue model can be decomposed to transform it into a series of semantically complete and logically coherent candidate reasoning steps. This decomposition not only focuses on the surface grammatical structure of the response but also analyzes the underlying reasoning logic and knowledge chain, ensuring that each step is an independent and understandable unit of reasoning.

[0219] Optionally, by converting candidate response information into candidate inference steps, the internal reasoning process of the dialogue model can be made explicit. That is, the thought process and decision-making process of the model when processing input information and generating response information can be presented in the form of text or structured data. The above method is crucial for subsequent inference alignment and semantic similarity evaluation, providing a clear framework that allows the inference paths of the dialogue model to be compared and evaluated.

[0220] Optionally, each candidate reasoning step should possess semantic integrity and logical coherence. This means that they independently express a key point of the answer, and that these key points, when combined, constitute a complete and reasonable answer. For example, in explaining a scientific concept, each reasoning step may involve definitions, principles, application examples, etc., collectively constructing a comprehensive understanding of the problem.

[0221] Step S308: Semantically align the candidate inference steps with the corresponding reference inference steps to obtain the alignment result. The reference inference step is used to represent the process by which the dialogue model analyzes the reference input information under the source language type. The semantic type of the reference inference information is different from the semantic type of the query information.

[0222] In this embodiment, semantic alignment of candidate reasoning steps and reference reasoning steps is a crucial step in the entire cross-language ability transfer method. It focuses on semantic alignment between candidate reasoning steps and source language reference reasoning steps to ensure that the response in the target language is consistent with the response in the high-resource language in terms of logic and information content.

[0223] Optionally, the purpose of semantic alignment is to evaluate and confirm the semantic consistency between candidate reasoning steps and reference reasoning steps, that is, to check whether the reasoning process of the target language response matches the reasoning process of the response in a high-resource language (such as English or Chinese). This is achieved by comparing key information, logical structure, and reasoning paths in responses from different languages.

[0224] Optionally, for semantic alignment, a cross-lingual sentence embedding model (such as LaBSE) is used to map candidate reasoning steps and reference reasoning steps to a unified high-dimensional semantic space, and cosine similarity between candidate reasoning steps and reference reasoning steps is calculated.

[0225] Optionally, during semantic alignment, an alignment matrix is ​​constructed that contains similarity scores for all pairwise combinations between candidate and reference inference steps. Each row of the alignment matrix represents a inference step in the candidate response, and each column represents a inference step in the source language reference response. Each element in the matrix is ​​the cosine similarity between the corresponding two-step inference, which helps to quantitatively assess the consistency of inference paths across different language responses.

[0226] Optionally, by semantically aligning each candidate reasoning step with the reference reasoning step, the alignment between the reasoning paths of the target language response and the source language response can be evaluated holistically. The average similarity between each candidate reasoning step and the reference reasoning step in the candidate response information can be calculated, or a suitable reasoning step matching method can be found to maximize the total similarity.

[0227] Step S310: Based on the alignment results, determine the target response information that matches the query information from at least one candidate response information.

[0228] In this embodiment, determining the target response information from the candidate response information based on the alignment result is a crucial stage in the cross-language ability transfer process. That is, the target language response information that matches the semantics of the source language question can be selected from the multiple candidate response information generated based on the alignment result.

[0229] Optionally, the similarity between candidate reasoning steps and source language reference reasoning steps is calculated and is typically represented as an alignment matrix. Each value in the matrix measures the semantic similarity between a certain reasoning step in the target language response and the corresponding step in the source language reference response; a higher value indicates a stronger semantic consistency between the two.

[0230] Optionally, the alignment results not only focus on the similarity between individual reasoning steps, but also evaluate the overall alignment between the candidate answer and the reference answer along the reasoning path. This can be achieved by calculating the average, weighted average, or other statistics of the non-zero elements in the alignment matrix, thus obtaining an index representing the consistency of the overall reasoning process.

[0231] Optionally, based on an assessment of overall alignment, the target language response that best matches the reasoning process of the source language reference response is selected from at least one candidate response. The selection process may employ a greedy algorithm, a global optimization algorithm, or other strategies to ensure that the selected response closely matches the source language response semantically and logically, while also being fluent and natural in the target language.

[0232] Optionally, if candidate responses involve multiple media formats (e.g., text, images, videos, etc.), multimodal alignment between the aforementioned different information sources can also be taken into account. This allows for additional evaluation mechanisms to ensure that the content of different modalities is consistent with the source language reference response as a whole, while presenting itself reasonably and naturally in the target language context.

[0233] Optionally, in scenarios with multiple candidate answers, the process of selecting the appropriate target response information based on the alignment results becomes more complex. The alignment scores of each candidate answer can be considered, along with other factors such as grammatical correctness, contextual relevance, and user preferences, to ensure that the final selected answer performs well across multiple dimensions.

[0234] In steps S302 to S310 of this embodiment, query information in the target language type is obtained; the query information is input into the dialogue model for analysis to obtain at least one candidate response information; at least one candidate inference step of the candidate response information is obtained; the candidate inference step is semantically aligned with the corresponding reference inference step to obtain an alignment result; based on the alignment result, the target response information matching the query information is determined from the at least one candidate response information. Through inference step alignment and semantic alignment mechanisms, the inference ability and output quality of the dialogue model are successfully improved in cross-language scenarios, thereby achieving the technical effect of reducing the limitations of the dialogue model in processing interactive information and solving the problem of the significant limitations of the dialogue model in processing interactive information.

[0235] The method described in this embodiment will be further described below.

[0236] As an optional implementation, the query information and the target response information are multimodal information. The types of multimodal information include at least one of the following: text information containing character information, video frame information containing frame image information, and audio information. The types of target response information include at least one of the following: text information, image information, video information, and voice information.

[0237] In this embodiment, treating query information and target response information as multimodal information in the context of cross-language ability transfer is an extended implementation method that enhances the dialogue model's ability to handle complex and diverse information.

[0238] Optionally, when the query contains multiple information types, such as text, images, audio, or video, the multilingual large model needs to possess cross-modal understanding capabilities. For example, if the query is a composite request containing both textual descriptions and images, the dialogue model must simultaneously understand the text content and the information conveyed in the images to accurately answer the question. This requires the dialogue model to not only process text but also parse and understand non-verbal information in images, audio, and video, such as objects in images, tone of voice in audio, and actions in video.

[0239] Optionally, the target response information can also be multimodal, including but not limited to text, images, video, and speech. This means that the dialogue model can not only generate text responses, but also generate image, video clips, or speech responses based on the nature of the question and its context, conveying information in a more intuitive and vivid way. For example, for a question about how to assemble furniture, the dialogue model might generate a response containing text instructions and images of assembly steps, or for a music request, the dialogue model might directly generate an audio clip as a response.

[0240] Optionally, when processing multimodal information, information across different modalities can be aligned to ensure that the response matches the query information in each modality. For example, if the query information contains both text and images, the text portion of the target response should be aligned with the text reference information in the source language, while the image portion should be aligned with the image reference information in the source language, ensuring the accuracy and consistency of information across all modalities.

[0241] In this embodiment, by processing multimodal query information and generating multimodal target response information, the multilingual big model not only enhances cross-language capabilities in a single modality but also expands its application scope in cross-language multimodal scenarios. This approach enables the model to better serve the diverse needs of global users, such as providing richly illustrated or audiovisual multilingual services in scenarios like education, entertainment, and remote assistance, thereby improving user experience and the efficiency of information delivery.

[0242] In summary, treating query information and target response information as implementation methods of multimodal information requires a large multilingual model to possess cross-modal understanding and generation capabilities, and to comprehensively process different types of information such as text, images, audio, and video. Through cross-modal alignment and comprehensive scoring selection, it is possible to ensure that response information from different modalities remains consistent in logic, semantics, and information content in a multilingual environment, thereby achieving higher-quality cross-language capability transfer and multimodal service provision.

[0243] This application also provides a method for processing interactive information from the perspective of human-computer interaction. Figure 4 This is a flowchart of another method for processing interactive information according to an embodiment of this application, such as... Figure 4 As shown, the method may include the following steps:

[0244] Step S402: In response to the input operation applied to the operation interface, display the input information on the operation interface.

[0245] This embodiment illustrates a key step in the interaction process between the user interface and the multilingual large model: the system responds to the user's input through the operation interface and displays the user's input information on the operation interface.

[0246] Optionally, users can perform input operations through the user interface, such as clicking a text box, touching a region of the screen, entering text using the keyboard, or uploading files (such as images, audio, or video). These input operations initiate the process of interacting with the multilingual large model, indicating that the user desires specific information or wants to complete a task.

[0247] Optionally, in response to user actions, the system captures user input information on the user interface. This includes identifying the type of input (text, image, audio, etc.), parsing the input content, and converting the input information into a format that the model can process. For example, for text input, the system needs to identify the language type of the text; for image input, the system may need to extract key information using image recognition technology to input it into the model for processing.

[0248] Optionally, after capturing user input, the input information can be displayed on the user interface to confirm its correctness and completeness. This not only provides users with an opportunity to verify their input and ensure its accuracy, but also enhances the user's interactive experience with the system, allowing them to intuitively see the information the system is processing. The displayed information may include the user's original input, a translation of the input (if cross-language input is involved), a summary of the input, or keywords, depending on the type of input information and the system design.

[0249] Step S404: In response to the information response operation applied to the operation interface, the target response information matching the input information is displayed on the operation interface. The target response information is determined from at least one candidate response information based on the alignment result. The alignment result is obtained by semantically aligning at least one candidate reasoning step with the corresponding reference reasoning step. The candidate reasoning step represents the process of the dialogue model analyzing the input information, and the reference reasoning step represents the process of the dialogue model analyzing the reference input information. The language type of the reference input information is different from the language type of the input information. The candidate response information is obtained by the dialogue model from the analysis of the input information.

[0250] This embodiment illustrates how to respond to a user's interaction request and display target response information that matches the input information.

[0251] Optionally, after the user completes input on the interface, a response action is triggered (such as clicking the "Get Answer" button) to begin processing the input information, calling a multilingual large model for analysis and response generation. This step marks the user's transition from the input stage to the response receiving stage, a crucial part of the user experience.

[0252] Optionally, the target response is selected from multiple candidate responses based on the alignment results of the reasoning process described above. These candidate responses are generated by the dialogue model after analyzing the input information, while the reference reasoning steps are derived from high-quality responses to the source language instructions. By calculating the semantic similarity between the candidate reasoning steps and the reference reasoning steps, the system can select the response that semantically matches the input information and has a consistent reasoning path as the target response.

[0253] Optionally, after selecting the target response information, the target response information is presented to the user and displayed on the user interface. The displayed content may include response text, related images, audio, or video, depending on the type of response generated by the model and the user's specific needs. In multimodal scenarios, it can also be ensured that the response content of each modality matches the user's input information and remains consistent in semantics and logic.

[0254] In steps S402 to S404 of this embodiment, in response to an input operation on the interface, the input information is displayed on the interface; in response to an information response operation on the interface, the target response information matching the input information is displayed on the interface. Through reasoning step alignment and semantic alignment mechanisms, the reasoning ability and output quality of the dialogue model are successfully improved in cross-language scenarios, thereby achieving the technical effect of reducing the limitations of the dialogue model in processing interactive information and solving the problem of the significant limitations of the dialogue model in processing interactive information.

[0255] The method described in this embodiment will be further described below.

[0256] As an optional implementation, step S402, in response to an input operation applied to the operation interface, displays input information on the operation interface, including: in response to an input operation applied to the operation interface, displaying initial input information of the source language type on the operation interface; in response to a selection operation applied to the operation interface, selecting a target language type from at least one language type; and in response to a translation operation applied to the operation interface, displaying input information of the target language type on the operation interface, wherein the input information of the target language type is translated from the input information of the source language type, the dialogue model is trained based at least on information samples of the target language type and information samples of the source language type, and the sample size of the information samples of the target language type is lower than the sample size of the information samples of the source language type.

[0257] In this embodiment, the interactive design of the user interface is particularly crucial in the application of a multilingual large-scale model, ensuring that users can easily switch between different languages ​​and obtain the corresponding information. This embodiment illustrates a user interface interaction method that supports multilingual input and translation.

[0258] Optionally, when a user performs an input operation on the interface, the system can respond to and display the initial information entered by the user in the source language (such as English, Chinese, etc.). These steps ensure that the user can see the complete content of their input, facilitating verification and confirmation.

[0259] Optionally, this embodiment provides a mechanism for users to select a target language from at least one language type to meet their needs in different language environments. Users can specify the target language through controls such as drop-down menus and language switching buttons on the interface. Once the target language is selected, users can perform a translation operation to translate the input information in the source language into the target language. The above functions can be implemented using pre-trained machine translation modules or integrated translation services to ensure the accuracy and fluency of the translation.

[0260] Optionally, since the sample size of information in the target language type is lower than that in the source language type, this indicates that the target language may be a low-resource language, meaning that there is relatively little target language data available for training the dialogue model. In this case, directly using the target language to train the model may lead to poor performance. Therefore, translating the input information from the source language into the target language can serve as an indirect method of augmenting training data, helping the model to better understand and generate responses in the target language.

[0261] Optionally, the dialogue model is trained on information samples from at least the target language and the source language, although the target language samples may be limited. By using cross-lingual reasoning alignment and translation samples, the dialogue model can transfer high-quality reasoning capabilities from the source language to the target language, thereby improving performance in low-resource language environments.

[0262] Optionally, the user interface design can also take into account the dynamic nature of multilingual environments, updating the displayed content in real time based on the user's language selection to ensure responsiveness and accuracy of information display when switching between different languages. Furthermore, the interface should support multilingual layouts and fonts, adapting to different writing habits.

[0263] Optionally, by allowing users to directly select and translate languages ​​on the user interface, user engagement and a sense of control in a multilingual environment are increased. Users can more flexibly use their preferred language for input and interaction, improving the overall user experience.

[0264] This application embodiment also provides a method for processing interactive information from the Software as a Service (SaaS) side. Figure 5 This is a flowchart of another method for processing interactive information according to an embodiment of this application, such as... Figure 5 As shown, the method may include the following steps:

[0265] Step S502: Obtain input information by calling the first interface, wherein the first interface includes a first parameter, and the parameter value of the first parameter includes the input information.

[0266] Step S504: Input the input information into the dialogue model for analysis to obtain at least one candidate response.

[0267] Step S506: Obtain at least one candidate inference step of candidate response information, wherein the candidate inference step is used to represent the process by which the dialogue model analyzes the input information.

[0268] Step S508: Semantically align the candidate inference steps with the corresponding reference inference steps to obtain the alignment result. The reference inference step is used to represent the process by which the dialogue model analyzes the reference input information. The language type of the reference input information is different from the language type of the input information.

[0269] Step S510: Based on the alignment results, determine the target response information that matches the input information from at least one candidate response information.

[0270] Step S512: Output the target response information by calling the second interface, wherein the second interface includes a second parameter, and the parameter value of the second parameter includes the target response information.

[0271] In steps S502 to S512 of this embodiment, input information is obtained by calling a first interface; the input information is input into the dialogue model for analysis to obtain at least one candidate response information; at least one candidate inference step of the candidate response information is obtained; the candidate inference step is semantically aligned with the corresponding reference inference step to obtain an alignment result; based on the alignment result, the target response information matching the input information is determined from the at least one candidate response information; and the target response information is output by calling a second interface. Through the inference step alignment and semantic alignment mechanism, the inference ability and output quality of the dialogue model are successfully improved in cross-language scenarios, thereby achieving the technical effect of reducing the limitations of the dialogue model in processing interactive information and solving the problem of the significant limitations of the dialogue model in processing interactive information.

[0272] According to an embodiment of this application, an embodiment of an interactive information processing system is also provided. Figure 6 This is a schematic diagram of an interactive information processing system according to an embodiment of this application, such as... Figure 6 As shown, the interactive information processing system 600 may include: an information input terminal 601, an information processing terminal 602, and an information output terminal 603.

[0273] Information input terminal 601 is used to acquire input information.

[0274] Information processing terminal 602 is used to input input information into a dialogue model for analysis to obtain at least one candidate response information; obtain at least one candidate inference step of the candidate response information, wherein the candidate inference step is used to represent the process of the dialogue model analyzing the input information; perform semantic alignment between the candidate inference step and the corresponding reference inference step to obtain an alignment result, wherein the reference inference step is used to represent the process of the dialogue model analyzing reference input information, the semantics of the reference input information are the same as the semantics of the input information, and the language type of the reference input information is different from the language type of the input information; based on the alignment result, determine the target response information that matches the input information from at least one candidate response information.

[0275] Information output terminal 603 is used to output the target response information.

[0276] This application provides an interactive information processing system 600. Input information is acquired through an information input terminal 601. The input information is then fed into a dialogue model for analysis through an information processing terminal 602 to obtain at least one candidate response; at least one candidate inference step is obtained from the candidate response; the candidate inference steps are semantically aligned with their corresponding reference inference steps to obtain an alignment result; based on the alignment result, a target response matching the input information is determined from the at least one candidate response. The target response is output through an information output terminal 603. This system achieves the technical effect of improving and reducing the limitations of dialogue models in processing interactive information, thus solving the technical problem of the limitations of dialogue models in processing interactive information.

[0277] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application, such as the data to be verified, are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0278] The embodiments of this application will be further explained below.

[0279] Currently, with the breakthroughs achieved by large language models in the field of natural language processing, the development of multilingual large models has also received widespread attention. However, existing multilingual large models face significant challenges in practical applications: compared to high-resource languages ​​such as Chinese and English, other languages ​​generally suffer from a scarcity of high-quality instruction data. This imbalance in training data leads to the model performing significantly worse on low-resource languages ​​than on high-resource languages ​​after instruction fine-tuning, resulting in a significant performance gap.

[0280] Against the backdrop of deepening globalization, the demand for cross-language interaction and multilingual services is rapidly increasing. This trend makes improving the versatility of large-scale multilingual models a research topic of significant theoretical and practical value. The versatility of large-scale multilingual models refers to their ability to understand and process language-independent knowledge in cross-language scenarios. This knowledge includes, but is not limited to, common-sense reasoning, logical analysis, and mathematical operations. The essence of these abilities is a cognitive process independent of specific language forms, and theoretically, they should be transferable and shareable across different languages.

[0281] Figure 7 This is a schematic diagram illustrating a common capability in related technologies: cross-language alignment. Figure 7As shown, when asked about the direction of the Coriolis force, regardless of whether the question is asked in Chinese or Japanese, the multilingual model should be able to accurately express the same scientific principle: it deflects to the right in the Northern Hemisphere and to the left in the Southern Hemisphere. The above example clearly demonstrates the consistency of language-independent knowledge across different languages, and also highlights the importance of multilingual models in maintaining knowledge consistency.

[0282] Therefore, how to effectively achieve cross-linguistic capability alignment and promote the effective transfer of knowledge between different languages ​​has become an urgent problem to be solved in the field of multilingual large model research.

[0283] In one related technology, Figure 8(a) is a schematic diagram of the literal translation method. As shown in Figure 8(a), the literal translation method directly converts the query (Chinese / English query) and response (Chinese / English response) of Chinese and English instruction data into the query (multilingual query) and response (multilingual response) of the target language through machine translation. Although the implementation is simple and direct, the following problems still exist: the quality of machine translation is unstable, and it is easy to introduce noise and errors; it is difficult to accurately convey the unique expressions and cultural connotations of the language. Therefore, translation errors will cause the model to learn inaccurate response patterns.

[0284] In another related technique, Figure 8(b) illustrates a translation-generation method. As shown in Figure 8(b), the translation-generation method translates Chinese / English query data into the target language (multilingual query), and then uses a large model to generate the corresponding multilingual response. Although the above method can maintain a certain degree of language nativeness, it faces the challenge that the generation path of the multilingual response differs from the original Chinese / English response. That is, the reasoning process between different languages ​​may be inconsistent, making it difficult to ensure semantic alignment of cross-language outputs. This not only leads to unsatisfactory alignment results in model training but may also impair the reasoning performance of high-resource languages ​​(Chinese / English).

[0285] In summary, the relevant technologies mainly suffer from two core problems: relying entirely on machine translation easily introduces errors, affecting training quality; differences in generation paths between different languages ​​lead to unsatisfactory cross-language alignment, impacting model performance in high-resource languages. Therefore, the technical challenge of dialogue models' limited processing of interactive information remains.

[0286] Furthermore, this application provides a cross-linguistic ability transfer method based on reasoning process alignment. By improving the consistency between multilingual responses and responses in high-resource terms (e.g., English and Chinese), the reasoning ability of the large model in English is transferred to multilingual scenarios, further enhancing the multilingual understanding and generation capabilities of the large model. This achieves the technical effect of reducing the limitations of dialogue models in processing interactive information, solving the problem of the significant limitations of dialogue models in processing interactive information.

[0287] The method described in this embodiment will be further described below.

[0288] Figure 9 This is a schematic diagram of a cross-language ability transfer process based on inference process alignment according to an embodiment of this application, as shown below. Figure 9 As shown, the user inputs a Chinese / English query (Chinese / English question), which is a query in a high-resource language, providing a benchmark for subsequent multilingual reasoning ability transfer. Machine translation technology is used to translate the Chinese or English question into multiple languages. The above steps can extend the question in the Chinese / English query to a multilingual space, preparing the foundation for subsequent multilingual response generation. The multilingual query is fed into a Multilingual Model (MLLM) to generate specific responses for each language. This is the stage where the Multilingual Model demonstrates its multilingual understanding and generation capabilities. For example, the Multilingual Model can generate three multilingual answers: Multilingual Answer 91, Multilingual Answer 92, and Multilingual Answer 93. Multilingual Answer 91 can generate reasoning step 91a based on its own length and structure. Multilingual Answer 92 can generate reasoning steps 92a, 92b, and 92c based on its own length and structure. Multilingual Answer 93 can generate reasoning steps 93a, 93b, and 93c based on its own length and structure. The Chinese / English answer 94 corresponding to the Chinese / English question can generate reasoning steps 94a, 94b, and 94c based on its length and structure. The reasoning steps split from the Chinese / English answer 94 and the multilingual answer can be aligned.

[0289] like Figure 9As shown, the "X" markings in the figure can represent potential problem points or failure points in the entire process. In the application embodiment, "X" can represent one of the following situations: If machine translation errors occur, the generated multilingual query may not match the original meaning, thereby affecting the accuracy and quality of subsequent multilingual responses. This is a problem that "X" can point to, emphasizing the importance of translation quality to the entire process. During the processing of multilingual responses, if the reasoning step of a certain response cannot be effectively aligned semantically with the step of the high-resource language response, the response may be marked as invalid or unsatisfactory, and "X" is used to represent such a failed alignment attempt. In reasoning step 91a, the initial screening of length or structure may not be able to effectively eliminate unreasonable candidate answers, resulting in the need to process noise or irrelevant answers in subsequent processing steps, and "X" is used to mark the potential loopholes in the above screening process.

[0290] In this embodiment, during the cross-language question conversion and answer generation process, the source language question (Chinese / English question) can be converted across languages ​​using a large language model to generate the target language question (multilingual question). The large model then generates k candidate answers (e.g., k=3, generating multilingual answers 91 / 92 / 93 and their corresponding Chinese / English answers). Through preliminary length comparison and filtering, candidate answers with significantly different lengths from the source language answers are eliminated (e.g., ...). Figure 9 (Multilingual responses in 91 languages) to ensure the structure of the responses is reasonable.

[0291] In this embodiment, during the decomposition and semantic representation of inference steps, the retained candidate answers (e.g., multilingual answers 92 and 93) and the source language answers can be decomposed according to sentence structure (e.g., newline characters) to extract key inference steps (inference steps 92a-92c and 93a-93c). The decomposition principle is based on sentence integrity, with each inference step corresponding to a complete inference unit. The cross-lingual sentence embedding model LaBSE is used to map the inference steps of different languages ​​to a unified high-dimensional semantic space, achieving cross-lingual semantic comparability through vectorized representation.

[0292] In this embodiment, during the answer selection process based on reasoning consistency, a step-level alignment matrix can be constructed, and the cosine similarity between the source language and the target language at each reasoning step can be calculated. The semantic alignment between steps is measured using cosine similarity, and the target language answer with the highest overall reasoning path score is selected as the final output.

[0293] In the embodiments of this application, the fine-grained reasoning step alignment method described above can effectively ensure that the multilingual output remains consistent with the high-resource language in the logical reasoning process, thereby realizing cross-language transfer of knowledge and skills.

[0294] Optionally, to address the quality issues caused by excessive reliance on machine translation, this application embodiment establishes a multi-candidate answer generation mechanism based on a large language model, significantly reducing the dependence on machine translation. Specifically: similar to the rejection sampling approach, by adjusting hyperparameters such as temperature (e.g., setting temperature = 0.7) and setting the hyperparameter k for the number of candidate answer samples (e.g., k = 8), multiple candidate answers are directly generated in the target language space using a large language model, constructing a richer answer space, increasing the probability of obtaining high-quality answers, while ensuring the naturalness of language expression and language-specific pragmatic features. Through a subsequent reasoning consistency screening mechanism, the impact of a single generation error on the final result is effectively reduced.

[0295] Optionally, to address the issue of inconsistent cross-language reasoning paths affecting high-resource linguistic capabilities, this application proposes a cross-language knowledge transfer method based on fine-grained reasoning steps. Specifically, answers in different languages ​​can be deconstructed into semantically complete sequences of reasoning steps, achieving a mapping from overall semantics to fine-grained reasoning processes. A pre-trained cross-language sentence embedding model, LaBSE, is used to map the reasoning steps of each language to a unified semantic space. By calculating the semantic similarity between reasoning steps, the consistency of answers in different languages ​​in the reasoning process is quantitatively evaluated. Candidate answers are selected based on the overall alignment of reasoning steps, ensuring the accuracy of cross-language knowledge transfer while maximizing the model's performance on high-resource languages.

[0296] In this embodiment, by decomposing and aligning the reasoning paths between different languages, knowledge transfer from high-resource languages ​​to low-resource languages ​​is realized in a unified semantic representation space. This effectively improves the generalization ability of the multilingual large model in cross-language scenarios, while maintaining the model's performance level on high-resource languages ​​to the maximum extent. This achieves the technical effect of reducing the limitations of dialogue models in processing interactive information and solves the problem of the large limitations of dialogue models in processing interactive information.

[0297] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0298] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0299] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, they can also be implemented by hardware. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application.

[0300] According to embodiments of this application, a method for implementing the above is also provided. Figure 2 The interactive information processing device shown is an interactive information processing method.

[0301] Figure 10 This is a schematic diagram of an interactive information processing device according to an embodiment of this application, such as... Figure 10 As shown, the interactive information processing device 1000 may include: a first acquisition unit 1002, a second acquisition unit 1004, a first alignment unit 1006, and a first determination unit 1008.

[0302] The first acquisition unit 1002 is used to acquire input information and input the input information into the dialogue model for analysis to obtain at least one candidate response information.

[0303] The second acquisition unit 1004 is used to acquire at least one candidate reasoning step of candidate response information.

[0304] The first alignment unit 1006 is used to perform semantic alignment between the candidate reasoning step and the corresponding reference reasoning step to obtain the alignment result.

[0305] The first determining unit 1008 is used to determine, based on the alignment result, the target response information that matches the input information from at least one candidate response information.

[0306] Here, the first acquisition unit 1002, the second acquisition unit 1004, the first alignment unit 1006, and the first determination unit 1008 correspond to steps S202 to S208 in the embodiments. The four units and the corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the above embodiments. It should be noted that the above units can be hardware components or software components stored in memory (e.g., memory 1404) and processed by one or more processors (e.g., processor 1402). The above units can also be part of a device and run in the computer terminal provided in the following embodiments.

[0307] According to embodiments of this application, a method for implementing the above is also provided. Figure 3 The interactive information processing device shown is an interactive information processing method.

[0308] Figure 11 This is a schematic diagram of another interactive information processing device according to an embodiment of this application, such as... Figure 11 As shown, the interactive information processing device 1100 may include: a third acquisition unit 1102, a first analysis unit 1104, a fourth acquisition unit 1106, a second alignment unit 1108, and a second determination unit 1110.

[0309] The third acquisition unit 1102 is used to acquire query information under the target language type.

[0310] The first analysis unit 1104 is used to input the query information into the dialogue model for analysis to obtain at least one candidate response information.

[0311] The fourth acquisition unit 1106 is used to acquire at least one candidate reasoning step of candidate response information.

[0312] The second alignment unit 1108 is used to perform semantic alignment between the candidate reasoning step and the corresponding reference reasoning step to obtain the alignment result.

[0313] The second determining unit 1110 is used to determine, based on the alignment result, the target response information that matches the query information from at least one candidate response information.

[0314] It should be noted that the third acquisition unit 1102, the first analysis unit 1104, the fourth acquisition unit 1106, the second alignment unit 1108, and the second determination unit 1110 mentioned above correspond to steps S302 to S310 in the above embodiments. The five units and the corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the above embodiments. It should be noted that the above units can be hardware components or software components stored in memory (e.g., memory 1404) and processed by one or more processors (e.g., processor 1402). The above units can also be part of a device and run in the computer terminal provided in the following embodiments.

[0315] According to embodiments of this application, a method for implementing the above is also provided. Figure 4 The interactive information processing device shown is an interactive information processing method.

[0316] Figure 12 This is a schematic diagram of another interactive information processing device according to an embodiment of this application, such as... Figure 12 As shown, the interactive information processing device 1200 may include: a first display unit 1202 and a second display unit 1204.

[0317] The first display unit 1202 is used to respond to input operations applied to the operation interface and display input information on the operation interface.

[0318] The second display unit 1204 is used to respond to information response operations on the operation interface and display target response information that matches the input information on the operation interface.

[0319] Here, the first display unit 1202 and the second display unit 1204 correspond to steps S402 to S404 in the above embodiments. The two units and the corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the above embodiments. It should be noted that the above units can be hardware components or software components stored in memory (e.g., memory 1404) and processed by one or more processors (e.g., processor 1402). The above units can also be part of a device and run in the computer terminal provided in the following embodiments.

[0320] According to embodiments of this application, a method for implementing the above is also provided. Figure 5 The interactive information processing device shown is an interactive information processing method.

[0321] Figure 13 This is a schematic diagram of another interactive information processing device according to an embodiment of this application, such as... Figure 13As shown, the interactive information processing device 1300 may include: a fifth acquisition unit 1302, a second analysis unit 1304, a sixth acquisition unit 1306, a third alignment unit 1308, a third determination unit 1310, and an output unit 1312.

[0322] The fifth acquisition unit 1302 is used to acquire input information by calling the first interface.

[0323] The second analysis unit 1304 is used to input the input information into the dialogue model for analysis to obtain at least one candidate response information.

[0324] The sixth acquisition unit 1306 is used to acquire at least one candidate reasoning step of candidate response information.

[0325] The third alignment unit 1308 is used to perform semantic alignment between the candidate reasoning step and the corresponding reference reasoning step to obtain the alignment result.

[0326] The third determining unit 1310 is used to determine, based on the alignment result, the target response information that matches the input information from at least one candidate response information.

[0327] Output unit 1312 is used to output target response information by calling the second interface.

[0328] Here, the fifth acquisition unit 1302, the second analysis unit 1304, the sixth acquisition unit 1306, the third alignment unit 1308, the third determination unit 1310, and the output unit 1312 correspond to steps S502 to S512 in the above embodiments. The six units and the corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the above embodiments. It should be noted that the above units can be hardware components or software components stored in memory (e.g., memory 1404) and processed by one or more processors (e.g., processor 1402). The above units can also be part of a device and run in the computer terminal provided in the following embodiments.

[0329] In this interactive information processing device, if a dialogue model is needed to respond to input information, the acquired input information can be input into the dialogue model. The dialogue model analyzes the input information to obtain candidate response information. At least one candidate reasoning step corresponding to the candidate response information can be obtained; that is, the process of analyzing / reasoning with the input information using the dialogue model to obtain the corresponding candidate response information can be acquired. Reference reasoning steps obtained by analyzing reference input information that has the same semantics as the input information but a different language type can be acquired, and the candidate reasoning steps and reference reasoning steps are semantically aligned to obtain an alignment result. Based on the alignment result, the target response information that can respond to the input information is determined from the candidate response information. In this embodiment, by semantically aligning with the reference reasoning steps, mature reasoning capabilities in high-resource languages ​​can be transferred to low-resource languages, significantly improving the universality and adaptability of the dialogue model in multilingual environments. Furthermore, by aligning the reasoning steps, consistency in the logical reasoning process of responses in different languages ​​is ensured, avoiding semantic errors or information distortion caused by language differences. This achieves the technical effect of reducing the limitations of dialogue models in processing interactive information and solves the problem of the limitations of dialogue models in processing interactive information.

[0330] It should be noted that the preferred embodiments involved in the above embodiments of this application are the same as the solutions, application scenarios and implementation processes provided in the above embodiments, but are not limited to the solutions provided in the above embodiments.

[0331] Embodiments of this application may provide a computing device. Figure 14 This is a structural block diagram of a computing device according to an embodiment of this application. Figure 14 As shown, the computing device 1400 may include one or more (only one is shown in the figure) processors 1402, memory 1404, a storage controller, and peripheral interfaces. The processor 1402 can be connected to a dialogue model to process interactive information. It should be noted that the dialogue model can be deployed in the cloud or on a terminal; no specific limitation is made here.

[0332] The aforementioned computing device can be understood as an integrated smart terminal, including but not limited to servers, desktop computers, PCs (Personal Computers), all-in-one model machines, etc., and the computing device may have the model in the above embodiments of this application pre-installed.

[0333] Specifically, this computing device can pre-install various types of models, including but not limited to models in natural language processing, visual processing, speech processing, code processing, and multimodal task processing, thus providing diverse model selection. In different product forms, this computing device can support one or more model usage methods, including but not limited to model training, model invocation, model fine-tuning, model deployment, model inference, and application. In some product forms, this computing device also supports model management, including but not limited to multi-type model management (supporting the management of discriminative, generative, and other model types), model version control (supporting the control of different model versions), and model evaluation (evaluating model performance and effectiveness based on model evaluation tools). In other product forms, this computing device can also create applications based on models, providing Application Programming Interface (API) invocation capabilities. Models can be invoked into created applications through the API interface, and application management tools are provided for application management and monitoring.

[0334] Furthermore, the computing device may also include data management (supporting the creation and management of model tuning datasets), a training center (providing abundant training resources to help users learn and master artificial intelligence (AI) technology), and basic control capabilities (providing enterprise-level basic control capabilities to ensure the security and efficient operation of the system). Through the above functions, it provides a comprehensive and integrated AI research and development, training, deployment, and application device.

[0335] The memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the methods and apparatus in the embodiments of this application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby implementing the methods in the above embodiments. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0336] The processor can invoke an executable program stored in memory via a transmission device to execute any of the methods described in the above embodiments.

[0337] Those skilled in the art will understand that, Figure 14The structure shown is for illustrative purposes only. The computing device can also be a smartphone, such as an Android phone or an iOS phone, or a tablet computer, PDA, mobile internet device (MID), personal digital assistant (PAD), or other terminal device. This diagram does not limit the structure of the computing device described above. For example, computing device 1400 may include more or fewer components (such as network interfaces, display devices, etc.) than shown in this diagram, or may have a different configuration than shown in this diagram.

[0338] Embodiments of this application may provide an electronic device. Figure 15 This is a structural block diagram of an electronic device according to an embodiment of this application. Figure 15 As shown, the electronic device may include: an input / output device 1512; a memory 1514; and a processor 1516, wherein the processor 1516 is connected to the input / output device 1512 and the memory 1514 via a bus 1518. The processor 1516 can be connected to a dialogue model for processing interactive information by invoking the dialogue model. It should be noted that the dialogue model can be deployed in the cloud or on a terminal; no specific limitation is made here.

[0339] The memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the methods and apparatus in the embodiments of this application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby implementing the methods in the above embodiments. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0340] The processor can invoke an executable program stored in memory via a transmission device to execute any of the methods described in the above embodiments.

[0341] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0342] Optionally, in this embodiment, the storage medium may be located in a computing device.

[0343] Optionally, in this embodiment, the computer-readable storage medium is configured to store an executable program, which, when the executable program is running, controls the device where the computer-readable storage medium is located to execute any of the methods described in the above embodiments.

[0344] Embodiments of this application also provide a computer program product. Optionally, in this embodiment, the computer program product may include a computer program that, when executed by a processor, implements the methods provided in the embodiments described above.

[0345] Embodiments of this application also provide a computer program product. Optionally, the computer program product may include a non-volatile computer-readable storage medium, which can be used to store a computer program that, when executed by a processor, implements the method provided in the above embodiments.

[0346] Embodiments of this application also provide a computer program. Optionally, in this embodiment, when the computer program is executed by a processor, it implements the method provided in the above embodiments.

[0347] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0348] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of units or modules may be electrical or other forms.

[0349] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0350] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0351] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory, random access memory, portable hard drive, magnetic disk, or optical disk.

[0352] The above are merely preferred embodiments of this application. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for processing interactive information, characterized in that, include: Obtain input information and input the input information into the dialogue model for analysis to obtain at least one candidate response. At least one candidate reasoning step for obtaining the candidate response information, wherein the candidate reasoning step is used to represent the process by which the dialogue model analyzes the input information; The candidate inference step and the corresponding reference inference step are semantically aligned to obtain the alignment result. The reference inference step is used to represent the process by which the dialogue model analyzes the reference input information. The language type of the reference input information is different from the language type of the input information. Based on the alignment result, a target response that matches the input information is determined from at least one of the candidate response information.

2. The method according to claim 1, characterized in that, The semantics of the reference input information are the same as the semantics of the input information.

3. The method according to claim 1, characterized in that, At least one candidate reasoning step for obtaining the candidate response information includes: Determine the structural information of the candidate response information, wherein the structural information of the candidate response information is used to represent the statement structure of the candidate response information; Based on the structural information of the candidate response information, the candidate response information is decomposed to obtain the candidate reasoning steps.

4. The method according to claim 3, characterized in that, Based on the structural information of the candidate response information, the candidate response information is decomposed to obtain the candidate reasoning steps, including: Based on the structural information of the candidate response information, the candidate response information is decomposed to obtain multiple reasoning steps, wherein the reasoning steps include the reasoning logic of the dialogue model in the process of analyzing the input information; The key reasoning steps among the multiple reasoning steps are identified as the candidate reasoning steps, wherein the importance of the key reasoning steps to the candidate response information is greater than an importance threshold.

5. The method according to claim 3, characterized in that, Determine the structure information of the candidate response information, including: Determine the information content of the candidate response information and the information content of the reference response information corresponding to the reference input information; Determine the difference in information content between the candidate response information and the reference response information; In response to the difference information being less than the information content threshold, the structural information of the candidate response information is determined.

6. The method according to claim 1, characterized in that, The language type of the input information is the target language type, the language type of the reference input information is the source language type, and the dialogue model is trained based on information samples under the target language type and information samples under the source language type. The sample size of the information samples under the target language type is lower than the sample size of the information samples under the source language type.

7. The method according to claim 6, characterized in that, The candidate reasoning steps are semantically aligned with their corresponding reference reasoning steps to obtain alignment results, including: The candidate inference steps under the target language type and the reference inference steps under the source language type are respectively mapped to the semantic space; The mapped candidate reasoning steps and the mapped reference reasoning steps are semantically aligned to obtain the alignment result.

8. The method according to claim 7, characterized in that, Mapping the candidate inference steps under the target language type and the reference inference steps under the source language type to the semantic space, respectively, includes: The candidate inference steps under the target language type and the reference inference steps under the source language type are respectively input into the information conversion model, wherein the information conversion model is obtained by training the encoder; Using the information conversion model, the candidate inference steps under the target language type and the reference inference steps under the source language type are respectively converted into vectors in the semantic space, wherein the vector dimension in the semantic space is greater than the dimension threshold.

9. The method according to claim 7, characterized in that, Semantic alignment is performed on the mapped candidate inference steps and the mapped reference inference steps to obtain the alignment result, including: The semantic similarity between the vectors of the candidate inference steps under the target language type in the semantic space and the vectors of the reference inference steps under the source language type in the semantic space is determined, wherein the dimension of the vectors in the semantic space is greater than a dimension threshold, and the semantic similarity is used to represent the alignment result.

10. The method according to claim 1, characterized in that, At least one of the candidate reasoning steps is used to construct a reasoning path for the dialogue model to analyze the input information. Based on the alignment result, a target response information matching the input information is determined from at least one of the candidate response information, including: Based on the semantic similarity corresponding to the candidate inference steps, the evaluation result of the inference path is determined, wherein the semantic similarity is used to represent the alignment result; Based on the evaluation results, the target response information is determined from at least one of the candidate response information.

11. The method according to any one of claims 1 to 10, characterized in that, Obtain input information, including: Obtain the initial input information under the source language type; The initial input information is converted from the source language type to the target language type to obtain the input information. The dialogue model is trained based on information samples in the target language type and information samples in the source language type. The sample size of the information samples in the target language type is lower than the sample size of the information samples in the source language type.

12. The method according to claim 11, further comprising: The initial input information under the source language type is determined as the reference input information; The reference input information is input into the dialogue model for analysis to obtain the reference response information under the source language type.

13. The method according to claim 12, further comprising: Determine the structural information of the reference response information, wherein the structural information of the reference response information is used to represent the statement structure of the reference response information; Based on the structural information of the reference response information, the reference response information is decomposed to obtain the reference reasoning steps.

14. The method according to any one of claims 1 to 10, characterized in that, The method further includes: In response to the reference adjustment operation, the parameters of the dialogue model are adjusted; The input information is input into the dialogue model for analysis to obtain at least one candidate response information, including: inputting the input information into the adjusted dialogue model to obtain a number of candidate response information corresponding to the adjusted parameters.

15. A method for processing interactive information, characterized in that, include: Obtain query information in the target language type, wherein the query information is used to represent information to be answered in the dialogue scenario; The query information is input into a dialogue model for analysis to obtain at least one candidate response. The dialogue model is trained based on information samples in the target language type and information samples in the source language type. The sample size of the information samples in the target language type is lower than the sample size of the information samples in the source language type. At least one candidate reasoning step for obtaining the candidate response information, wherein the candidate reasoning step is used to represent the process by which the dialogue model analyzes the input information; The candidate reasoning step and the corresponding reference reasoning step are semantically aligned to obtain the alignment result. The reference reasoning step is used to represent the process by which the dialogue model analyzes the reference input information under the source language type. The language type of the reference reasoning information is different from the language type of the query information. Based on the alignment result, a target response that matches the query information is determined from at least one of the candidate response information.

16. The method according to claim 15, characterized in that, The inquiry information and the target response information are multimodal information. The types of multimodal information include at least one of the following: text information containing character information, video frame information containing frame image information, and audio information. The types of target response information include at least one of the following: text information, image information, video information, and voice information.

17. A method for processing interactive information, characterized in that, include: In response to an input operation applied to the user interface, the input information is displayed on the user interface. In response to an information response operation applied to the operation interface, a target response information matching the input information is displayed on the operation interface. The target response information is determined from at least one candidate response information based on an alignment result. The alignment result is obtained by semantically aligning at least one candidate inference step with its corresponding reference inference step. The candidate inference step represents the process by which the dialogue model analyzes the input information, and the reference inference step represents the process by which the dialogue model analyzes reference input information. The language type of the reference input information is different from the language type of the input information. The candidate response information is obtained by the dialogue model from the analysis of the input information.

18. A method for processing interactive information, characterized in that, include: Input information is obtained by calling a first interface, wherein the first interface includes a first parameter, and the parameter value of the first parameter includes the input information; The input information is fed into the dialogue model for analysis to obtain at least one candidate response. At least one candidate reasoning step for obtaining the candidate response information, wherein the candidate reasoning step is used to represent the process by which the dialogue model analyzes the input information; The candidate inference step and the corresponding reference inference step are semantically aligned to obtain the alignment result. The reference inference step is used to represent the process by which the dialogue model analyzes the reference input information. The language type of the reference input information is different from the language type of the input information. Based on the alignment result, a target response that matches the input information is determined from at least one of the candidate response information; The target response information is output by calling a second interface, wherein the second interface includes a second parameter, and the parameter value of the second parameter includes the target response information.

19. A computing device, characterized in that, include: Memory, which stores executable programs; A processor for running the program, wherein the program, when running, performs the method according to any one of claims 1 to 18.

20. An electronic device, characterized in that, include: Memory, which stores executable programs; A processor, connected to the memory via a bus, is used to run the program, wherein the program, when running, executes the method according to any one of claims 1 to 18.

21. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored executable program, wherein, when the executable program is executed, it controls the device on which the computer-readable storage medium is located to perform the method according to any one of claims 1 to 18.

22. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the method described in any one of claims 1 to 18.

Images