Multi-language large model optimization training method, electronic equipment, storage medium and program product

Abstract

The invention discloses a multi-language large model optimization training method, electronic equipment, a storage medium and a program product. The method comprises the steps of obtaining a multi-language data set of multiple languages supported by a first initial multi-language large model; inputting at least one kind of language data in the multi-language data set into a first initial multi-language large model to determine a neuron set corresponding to an optimization training target; performing preset processing on neuron parameters corresponding to the neuron set according to an optimization training target to obtain a second initial multi-language large model; and training the second initial multi-language large model by adopting a training sample set corresponding to the optimization training target. The multi-language large model is optimized and trained in combination with the neuron set corresponding to the optimization training target, the problem of disastrous forgetting in the prior art is solved, and the training efficiency of the model is improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a multi-language large model optimization training method, electronic device, storage medium and program product. Background Art

[0002] In current multilingual large language models (e.g., BLOOM, LLaMA, etc.), when trying to add a new language to the model, the "catastrophic forgetting" problem is often encountered, that is, the model will forget the knowledge of the existing language while learning the new language. This is because the model will update the weights of the entire network when retraining, causing the previously learned knowledge to be overwritten. Therefore, a technical solution that can solve the "catastrophic forgetting" problem in multilingual large language models is urgently needed. Summary of the invention

[0003] The embodiments of the present application provide a multi-language large model optimization training method, electronic device, storage medium and program product, which are used to solve at least one of the above-mentioned technical problems.

[0004] In a first aspect, an embodiment of the present application provides a multi-language large model optimization training method, comprising: Acquire a multilingual dataset of multiple languages ​​supported by the first initial multilingual large model; Inputting at least one language data in the multilingual data set into the first initial multilingual large model to determine a neuron set corresponding to an optimized training target; Presetting the neuron parameters corresponding to the neuron set according to the optimization training objective to obtain a second initial multilingual large model; The second initial multilingual large model is trained using a training sample set corresponding to the optimization training objective.

[0005] In some embodiments, the optimization training target includes adding new language processing capabilities to the multilingual large model; the neuron set corresponding to the optimization training target includes the key neuron set corresponding to the multiple languages ​​supported by the first initial multilingual large model.

[0006] In some embodiments, the neuron parameters corresponding to the neuron set are preset according to the optimization training objective to obtain a second initial multilingual large model, including: Freeze corresponding upstream and downstream weight parameters of key neuron sets corresponding to multiple languages ​​supported by the first initial multilingual large model to obtain a second initial multilingual large model.

[0007] In some embodiments, the optimization training goal includes enhancing the processing capability of the multilingual large model for the old languages ​​that are already supported; the neuron set corresponding to the optimization training goal includes the key neuron set corresponding to the old languages ​​supported by the first initial multilingual large model.

[0008] In some embodiments, the neuron parameters corresponding to the neuron set are preset according to the optimization training objective to obtain a second initial multilingual large model, including: Freeze the corresponding upstream and downstream weight parameters of the key neuron set corresponding to the old language supported by the first initial multilingual large model to obtain a second initial multilingual large model.

[0009] In some embodiments, the neuron parameters corresponding to the neuron set are preset according to the optimization training objective to obtain a second initial multilingual large model, including: Freeze the corresponding upstream and downstream weight parameters of the key neuron set corresponding to the languages ​​other than the old languages ​​supported by the first initial multilingual large model to obtain a second initial multilingual large model.

[0010] In some embodiments, the optimization training target includes forgetting the processing capability of the large multilingual model for the supported old languages; the neuron set corresponding to the optimization training target includes the key neuron set corresponding to the old languages ​​that need to be forgotten in the first initial large multilingual model.

[0011] In some embodiments, the neuron parameters corresponding to the neuron set are preset according to the optimization training objective to obtain a second initial multilingual large model, including: Reset corresponding upstream and downstream weight parameters of the key neuron set corresponding to the old language that needs to be forgotten in the first initial multilingual large model to zero to obtain a second initial multilingual large model.

[0012] In a second aspect, an embodiment of the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that the processor executes the computer program to implement the steps of the multi-language large model optimization training method described in any embodiment of the present application.

[0013] In a third aspect, an embodiment of the present application provides a computer-readable storage medium on which a computer program / instruction is stored, characterized in that when the computer program / instruction is executed by a processor, the steps of the multi-language large model optimization training method described in any embodiment of the present application are implemented.

[0014] In a fourth aspect, an embodiment of the present application provides a computer program product, including a computer program / instructions, characterized in that when the computer program / instructions are executed by a processor, the steps of the multi-language large model optimization training method described in any embodiment of the present application are implemented.

[0015] When optimizing the training of the multilingual large model, the present application first determines the neuron set in the first initial multilingual large model corresponding to the optimization training target, and then performs preset processing on the corresponding neuron parameters in the first initial multilingual large model according to the optimization training target, and finally uses the training sample set to train the multilingual large model. Through the preprocessing step, the influence on the neuron parameters corresponding to the preprocessing is avoided or reduced when the multilingual large model is trained using the training sample set, thereby solving the problem of "catastrophic forgetting" existing in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 A flowchart of an embodiment of the multi-language large model optimization training method of the present application; Figure 2 A schematic diagram of the structure of an embodiment of a computer device of the present application; Figure 3 A schematic diagram of the principle of LLMs in this application encoding input into "Lingua Franca" and then decoding "Lingua Franca" into the target language; Figure 4 Similarity maps of all sample pairs generated by the BLOOM-7b1 model in this application; Figure 5a Schematic diagram of the changes of language areas in this application along with the BLOOM-7b1 training steps; Figure 5b Schematic diagram of the change of semantic alignment in this application along with the BLOOM7-b1 training steps; Figure 6 A schematic diagram showing the distribution of key neurons in different layers of BLOOM-7b1 in this application; Figure 7 The distribution of key neurons in different languages ​​on BLOOM-7b1 in this application; Figure 8 A schematic diagram of the measurements of different size models across the BLOOM family in this application; Figure 9a is the similarity map of the LLaMA2-7b model in this application; Figure 9b is the similarity map of the LLaMA2-13b model in this application; Fig.9c is the similarity map of the LLaMA2-70b model in this application; Fig.10 It is a schematic diagram of the measurement of different size models of the LLaMA2 series in this application; Fig.11 Schematic diagram of the distribution of LLaMA2-7b key neurons in different languages ​​in this application. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0019] It should also be noted that, in this article, the terms "include" and "comprise" include not only those elements, but also other elements not explicitly listed, or elements inherent to such processes, methods, articles or equipment. In the absence of further restrictions, the elements defined by the sentence "include..." do not exclude the existence of other identical elements in the process, method, article or equipment that includes the elements.

[0020] In the process of exploring the technical solution of this application, the inventors have also tried the following solutions: Use unstructured pruning technology to find the most critical weights for a certain language and freeze these weights during training. Specifically, the method of defining the importance of weights in technologies such as Wanda and SparseGPT can be used to select the most important weights and keep them unchanged. However, the inventors found that finding unstructured weights can more accurately locate which weights are more important for a specific language, but there is no way to achieve training acceleration on current general-purpose GPUs, and the computational cost is equivalent to full model training.

[0021] like Figure 1As shown, an embodiment of the present application provides a multi-language large model optimization training method, the optimization training includes but is not limited to performing new language optimization training, old language enhancement training and old language forgetting training on the multi-language large model, the method includes: S10: Acquire multilingual datasets in multiple languages ​​supported by the first initial multilingual large model.

[0022] Exemplarily, the first initial multilingual large model can be a large language model such as BLOOM, LLaMA, etc. in the related art. The first initial multilingual large model already has the ability to process multiple languages ​​and is trained based on multilingual data sets of the multiple languages. It should be noted that in this application, the existing large language models such as BLOOM, LLaMA, etc. in the prior art can be directly used as the first initial multilingual large model, and the specific implementation method of pre-training the first initial multilingual large model is not limited.

[0023] S20, inputting at least one language data in the multilingual data set into the first initial multilingual large model to determine a neuron set corresponding to an optimized training objective.

[0024] Exemplarily, at least one language data in the multilingual data set is a parallel corpus of a language to be protected, and these parallel corpora are input into the first initial multilingual large model, the model activation state is obtained, and a neuron set corresponding to the optimization training target is calculated. The neurons in the neuron set include but are not limited to the output of the intermediate layer of the model, the rows or columns of the matrix involved in the calculation in the model, or the single elements in the matrix involved in the calculation in the model, etc.

[0025] S30. Preset the neuron parameters corresponding to the neuron set according to the optimization training target to obtain a second initial multilingual large model.

[0026] Exemplarily, the neuron parameters corresponding to the neuron set include the upstream and downstream weights of the neurons in the first initial multilingual large model connecting the neurons in the neuron set. The preset processing includes but is not limited to parameter freezing processing or fine-tuning or continued training.

[0027] S40: training the second initial multilingual large model using the training sample set corresponding to the optimization training objective.

[0028] When optimizing the training of the multilingual large model, the present application first determines the neuron set in the first initial multilingual large model corresponding to the optimization training target, and then performs preset processing on the corresponding neuron parameters in the first initial multilingual large model according to the optimization training target, and finally uses the training sample set to train the multilingual large model. Through the preprocessing step, the influence on the neuron parameters corresponding to the preprocessing is avoided or reduced when the multilingual large model is trained using the training sample set, thereby solving the problem of "catastrophic forgetting" existing in the prior art.

[0029] In some embodiments, optimizing the training target includes adding new language processing capabilities to the multilingual large model; the neuron set corresponding to the optimized training target includes the key neuron set corresponding to the multiple languages ​​supported by the first initial multilingual large model.

[0030] For example, a preset method (for details, see the relevant content of "3.3, Key Language Area Detection" in the subsequent embodiments of this application) can be used to search for the key area of ​​the original language of the model. The general steps are to input the parallel corpus of all the original languages ​​that need to be protected, obtain the model activation state, and calculate the neuron set that is key to all the original languages ​​of the model according to the algorithm.

[0031] In some embodiments, the neuron parameters corresponding to the neuron set are preset according to the optimization training target to obtain a second initial multilingual large model, including: freezing the corresponding upstream and downstream weight parameters of the key neuron set corresponding to the multiple languages ​​supported by the first initial multilingual large model to obtain the second initial multilingual large model. Furthermore, it also includes training the model using the corpus of the new language.

[0032] For example, during training, the upstream and downstream weights connecting the neurons in these key neuron sets are frozen. Specifically speaking, the MLP (Multi-layer Perceptron) structure consists of fully connected neurons with nonlinear activation functions, divided into at least three layers, and its notable feature is the ability to distinguish data that cannot be linearly separated. In BLOOM, these operations can be expressed as follows: z=W1x+b1(1) h = σ(z) (2) y=W2h+b2(3) Where x is the input vector, W i is the weight matrix, b i is the bias vector, h is the hidden layer output, y is the final output, and σ(.) is the activation function. In BLOOM, the activation function is GeLU.

[0033] For a neuron in the set hk , we will freeze the first k Column, and the first k Column, that is, the column with which the computation occurs for this neuron.

[0034] For other structures, after finding the key neurons, we can also freeze the elements that are calculated with these key neurons. In specific operations, the gradients flowing through these neurons can be set to zero.

[0035] In some embodiments, the optimization training goal includes enhancing the processing capability of the multilingual large model for the old languages ​​that are already supported; the neuron set corresponding to the optimization training goal includes the key neuron set corresponding to the old languages ​​supported by the first initial multilingual large model.

[0036] For example, a preset method (for details, see the relevant content of "3.3, Key Language Area Detection" in the subsequent embodiments of this application) can be used to search for the key area of ​​the original language of the model. The general steps are to input parallel corpora of all original languages ​​that need to be protected, obtain the model activation state, and calculate the set of neurons that are key to the language that needs to be enhanced according to the algorithm.

[0037] In an optional implementation, the language data corresponding to the old language that needs to be enhanced in the multilingual data set is input into the model (for example, the first initial multilingual large model), and the neurons in the model whose contribution score to the old language is greater than a certain contribution threshold are determined as key neurons. Among them, the contribution threshold can be set as needed, and this application does not limit it. For the specific algorithm for calculating the contribution score of neurons to the old language, please refer to the relevant content of "3.3, Key Language Area Detection" in the subsequent embodiments of this application).

[0038] In some embodiments, the neuron parameters corresponding to the neuron set are preset according to the optimization training target to obtain a second initial multilingual large model, including: freezing the corresponding upstream and downstream weight parameters of the key neuron set corresponding to the old language supported by the first initial multilingual large model to obtain the second initial multilingual large model. Further, it also includes training the model using the corpus of the language to be enhanced.

[0039] This embodiment can enhance the performance of a certain language of the model in a targeted manner, improve the efficiency of training, and reduce the resources consumed by training. When the model can add new languages ​​without sacrificing existing languages, the multi-language large model will become more flexible and scalable. This not only reduces the difficulty of maintaining a multi-language system, but also helps to expand the coverage of languages, especially in applications that need to continuously add new languages ​​or tasks, such as real-time translation systems or globalized dialogue agent systems.

[0040] For example, during training, the upstream and downstream weights connecting the neurons in these key neuron sets are frozen. Specifically for the MLP structure, for a neuron in the set h k , We will freeze the first k Column, and the first k Column, that is, the column where the calculation occurs with this neuron. For other structures, after finding the key neurons, we can also freeze the elements where the calculation occurs with these key neurons. In specific operations, the gradients flowing through these neurons can be set to zero.

[0041] In this embodiment, by freezing the upstream and downstream weight parameters of the key neurons related to the old language that needs to be enhanced, more new neurons that serve the old language that needs to be enhanced can be trained during the model training process, so that the trained model has more key neurons for the language that needs to be enhanced, so as to obtain a stronger processing capability for the language that needs to be enhanced.

[0042] In some embodiments, the neuron parameters corresponding to the neuron set are preset according to the optimization training objective to obtain a second initial multilingual large model, including: Freeze the corresponding upstream and downstream weight parameters of the key neuron set corresponding to the languages ​​other than the old languages ​​supported by the first initial multilingual large model to obtain a second initial multilingual large model. Further, it also includes training the model using the corpus of the language to be enhanced.

[0043] In this embodiment, by freezing the weight parameters related to neurons of languages ​​other than the old language that needs to be enhanced, the weight parameters of the upstream and downstream key neurons of the old language that needs to be enhanced can be more targeted during training, thereby improving the targetedness and training efficiency of the optimization training and reducing the resources consumed by training.

[0044] In some embodiments, the optimization training target includes forgetting the processing capability of the large multilingual model for the supported old languages; the neuron set corresponding to the optimization training target includes the key neuron set corresponding to the old languages ​​that need to be forgotten in the first initial large multilingual model.

[0045] For example, a preset method (for details, see the relevant content of "3.3, Key Language Area Detection" in the subsequent embodiments of this application) can be used to search for the key area of ​​the original language of the model. The general steps are to input parallel corpora of all original languages ​​that need to be protected, obtain the model activation state, and calculate the set of neurons that are key to the language that needs to be forgotten according to the algorithm.

[0046] In some embodiments, according to the optimization training target, the neuron parameters corresponding to the neuron set are preset to obtain a second initial multilingual large model, including: setting the corresponding upstream and downstream weight parameters of the key neuron set corresponding to the old language that needs to be forgotten in the first initial multilingual large model to zero to obtain the second initial multilingual large model. Furthermore, it also includes using corpora of other languages ​​that do not need to be forgotten to perform model training.

[0047] This application implements the model's ability to forget in a targeted way, which means that future multilingual models or multitask models can dynamically adjust themselves according to demand. In some scenarios, such as when a customized dialogue agent is required, the model can temporarily "forget" some unnecessary tasks or knowledge and focus on specific tasks. This flexibility can significantly improve the response efficiency of intelligent systems.

[0048] For example, the upstream and downstream weights of the neurons in these key neuron sets are set to zero. Specifically for the MLP structure, for a neuron in the set h k , we will set zero the first k Column, and the first k Column, that is, the column where the calculation occurs with this neuron. For other structures, after finding the key neurons, similarly, we can also set the elements calculated with these key neurons to zero.

[0049] In some embodiments, an embodiment of the present application provides a computer device, including a memory, a processor, and a computer program stored in the memory, characterized in that the processor executes the computer program to implement the steps of the multi-language large model optimization training method described in any embodiment of the present application.

[0050] In some embodiments, an embodiment of the present application provides a computer-readable storage medium on which a computer program / instruction is stored, characterized in that when the computer program / instruction is executed by a processor, the steps of the multi-language large model optimization training method described in any embodiment of the present application are implemented.

[0051] In some embodiments, the embodiments of the present application provide a computer program product, including a computer program / instructions, characterized in that when the computer program / instructions are executed by a processor, the steps of the multi-language large model optimization training method described in any embodiment of the present application are implemented.

[0052] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of actions combined, but those skilled in the art should be aware that the present application is not limited by the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application. In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0053] This application first determines the key neuron areas associated with each language, which are mainly responsible for processing information in a specific language. By freezing the weights of the upstream and downstream neurons related to the old language, the model can still retain the processing ability of the old language when training new tasks or new languages, avoiding forgetting the old tasks in the process of learning new tasks. Strengthening these weights through targeted training can also enhance the model's ability to learn a certain language, and improve the model's multilingual performance more efficiently. In addition, by freezing specific weights or targeted forgetting techniques, model training becomes more efficient, reducing the waste of resources from repeated full model training. This is especially important in environments with large amounts of data and tight computing resources. It can not only reduce training time, but also reduce energy consumption, thereby achieving more environmentally friendly AI applications.

[0054] Figure 2 is a schematic diagram of the hardware structure of a computer device for executing a multi-language large model optimization training method provided by another embodiment of the present application, such as Figure 2 As shown, the device includes: One or more processors 210 and memory 220, Figure 2 A processor 210 is taken as an example.

[0055] The device for executing the multi-language large model optimization training method may further include: an input device 230 and an output device 240.

[0056] The processor 210, the memory 220, the input device 230 and the output device 240 may be connected via a bus or other means. Figure 2 The example of connecting through bus is taken in the following.

[0057] The memory 220, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer executable programs and modules, such as program instructions / modules corresponding to the multi-language large model optimization training method in the embodiment of the present application. The processor 210 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions and modules stored in the memory 220, that is, the multi-language large model optimization training method in the above method embodiment is implemented.

[0058] The memory 220 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of the multi-language large model optimization training device, etc. In addition, the memory 220 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 220 may optionally include a memory remotely arranged relative to the processor 210, and these remote memories may be connected to the multi-language large model optimization training device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0059] The input device 230 can receive input digital or character information and generate signals related to user settings and function control of the multi-language large model optimization training device. The output device 240 can include display devices such as display screens.

[0060] The one or more modules are stored in the memory 220, and when executed by the one or more processors 210, the multi-language large model optimization training method in any of the above method embodiments is executed.

[0061] The above-mentioned product can execute the method provided in the embodiment of the present application, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not fully described in this embodiment, please refer to the method provided in the embodiment of the present application.

[0062] In order to make the technical solution and effect of this application clearer, the research, exploration, implementation and experimental demonstration process of the invention are presented as follows: Large language models (LLMs) have demonstrated remarkable performance in multilingual settings. While recent studies have shown that LLMs are able to transfer skills learned in one language to other languages, the internal mechanisms behind this remain unclear. We observe that LLMs show similar neuronal activation patterns when processing the same language, revealing the existence and location of key language regions. Furthermore, we find that neuronal activation patterns are similar when processing sentences with the same semantics in different languages. This suggests that LLMs map semantically identical inputs across languages ​​to a “lingua franca” (a common semantic latent space that allows consistent processing across languages). This semantic alignment becomes more pronounced with increasing training and model size, leading to more language-independent activation patterns. Furthermore, we find that key language neurons are concentrated in the first and last layers of LLMs, and that languages ​​with unique features as well as low-resource languages ​​are more challenging to align in the semantic space. Experiments based on BLOOM and LLaMA2 support these findings, highlighting the evolution of the structure of multilingual LLMs during training and scaling. This application provides insights into the inner workings of LLMs, laying the foundation for future improvements in their cross-lingual capabilities.

[0063] 1. Introduction In recent years, large language models (LLMs) have attracted much attention due to their remarkable performance. The multilingual capabilities of LLMs are a key research area, especially as AI technologies become more accessible to people with different backgrounds and native languages. Interestingly, recent studies have shown that LLMs can develop cross-lingual capabilities, transferring skills learned in one language to other untrained languages. Various methods have been explored to enhance the cross-lingual capabilities of multilingual LLMs, including instruction adjustment, forced alignment, and contrastive learning. However, the internal mechanisms of how multilingual LLMs operate and develop cross-lingual capabilities remain an under-researched topic. In this application, we will delve into the internal mechanisms of multilingual LLMs and explore how their internal structure changes with the training process and model size.

[0064] like Figure 3 As shown in the figure, it is a schematic diagram of the principle that LLMs in this application encode the input into "Lingua Franca" (i.e. "universal language", a latent semantic space representation shared by all languages), and then decode "Lingua Franca" into the target language. As shown in the figure, as the training progresses and the model scales up, the ability of LLMs to map the input to this universal semantic space will be improved.

[0065] In the field of neuroscience, studies have shown that when multilinguals process different languages, their brain’s language networks show different response patterns. It is believed that different language abilities are stored in different areas of the human brain. At the same time, people who are proficient in different languages ​​can perform the same task in different languages. The human brain shows similar activation patterns when processing the same task in different languages. Studies have also shown that as individuals become more proficient in a new language, the activation patterns in that language become more and more similar to those in other languages. This raises the question: do these phenomena also occur in LLMs? The answer is yes. We found that LLMs use different neurons to process inputs in different languages, but inputs with the same semantics in different languages ​​are mapped to a “lingua franca” (a common semantic latent space that allows consistent processing across languages).

[0066] First, by studying the activation of neurons in LLMs, we found that neuronal activations exhibited similar patterns when processing inputs of the same language. Certain key parameters in LLMs correspond to language abilities, and there are language-specific parameters for different languages. Our study supports this idea. By probing the neurons that contribute most to this similarity, we can identify the key language areas for a specific language. When the key language areas for a specific language are turned off, LLMs lose a significant amount of processing ability for that language, but the processing ability for other languages ​​remains unchanged.

[0067] Second, we found that the activation patterns of LLMs neurons were similar when processing inputs with the same semantics but in different languages. This suggests that multilingual LLMs map inputs to a common latent semantic space, enabling similar processing of information across languages ​​and facilitating cross-lingual ability transfer. We call this phenomenon semantic alignment.

[0068] Third, we find that as training progresses, the size of the critical language regions becomes smaller, while the semantic alignment becomes more distinct. Similarly, as the model size increases, the activation patterns of neurons become less affected by language, but the semantic alignment becomes more distinct. We define some metrics to facilitate the comparison of the differences in language regions and semantic alignment.

[0069] By studying the internal structure of LLMs, we found that neurons in key language regions are generally located in the first few layers and the last few layers. As training progresses and the model size increases, these key regions become more dense in the first few layers and the last few layers. Based on this information, we hypothesize that LLMs encode inputs into "Lingua Franca" (a latent semantic space representation shared by all languages) and then decode this "Lingua Franca" into the target language. As training or model size increases, the model can more effectively map inputs with the same semantics to a common semantic space and then perform reasoning.

[0070] We also found an interesting pattern: languages ​​with less similar linguistic features, such as languages ​​from different language families or with different writing systems, are generally more difficult to align to a common semantic space. For example, the critical language regions for Chinese and Arabic, which use different alphabets, are relatively large compared to Romance languages. This suggests that the model needs to spend more effort in adjusting these languages ​​to a common semantic space. In addition, poorly trained languages ​​(such as low-resource languages) align poorly, such as Vietnamese.

[0071] The experiments are mainly conducted on BLOOM (Workshop et al., 2023), one of the most effective open-source multilingual LLM models, using its published intermediate checkpoints to examine its evolution during training. To show the scalability of our results on other models, we also obtain similar results on LLaMa2.

[0072] 2. Background 2.1 Large Language Models (LLMs) and Multilingual LLMs Large language models such as GPT-4, LLaMA, and OPT have revolutionized natural language processing with their ability to understand and generate nuanced text. In addition, multilingual large models such as BLOOM and XLM-R overcome language barriers by learning universal representations from text in multiple languages. These models typically use a transformer architecture with only a decoder, and each layer consists of an attention module and a multi-layer perceptron module. Multilingual large language models usually incorporate multilingual data in the pre-training phase for better alignment.

[0073] The attention module is crucial in the transformer model, allowing the model to focus on different parts of the input sequence and assign different importance to each token. It enables the model to dynamically prioritize important information in the input and enhances the model's ability to capture dependencies and relationships, thereby improving performance.

[0074] 2.2 Multilayer Perceptron (MLP) A multilayer perceptron (MLP) consists of fully connected neurons with nonlinear activation functions, divided into at least three layers, and its notable feature is its ability to distinguish data that is not linearly separable. In BLOOM, these operations can be expressed as follows: z=W1x+b1(1) h = σ(z) (2) y=W2h+b2(3) Where x is the input vector, W i is the weight matrix, b i is the bias vector, h is the hidden layer output, y is the final output, and σ(.) is the activation function. In BLOOM, the activation function is GeLU.

[0075] In LLaMA, the MLP module can be expressed as: h1=W1x + b1(4) h2=W2x + b2(5) SwiGLU(x) = Swish(h1)⊙h2(6) y = W3・SwiGLU(x) + b3(7) Among them, the Swish function is a nonlinear activation function in deep learning.

[0076] As pointed out in the related art, MLP weights can store complex syntactic and semantic patterns, which are the building blocks of language. Therefore, in the following text, we refer to the neurons of BLOOM as h, which stands for the hidden layer output, and the neurons of LLaMA as the output of SwiGLU(x).

[0077] 3. Methods 3.1. Measuring the Similarity of Neuron Activation To separate the key language regions from the universal language (represented in the latent semantic space shared by all languages), we use the parallel corpus to activate the neurons of the LLMs. When the LLMs process each token, we record the neuron activation results (e.g., the hidden layer output h of the MLP layer in BLOOM or the output of SwiGLU(x) in LLaMA). We then average these results over all tokens in each sample. By concatenating the average results of all layers, we create an activation vector for each sample and normalize it.

[0078] Finally, we calculate the cosine similarity between each pair of samples and generate a similarity graph.

[0079] Neuron activation extraction:

[0080] in, Representation sample s i In layer m Middle pair mark t Activation, T si It is a sample s i The number of markers in . s i The number of tags in .

[0081] Concatenation of layer activations:

[0082] Here, M is the total number of layers in the model.

[0083] Cosine similarity plot:

[0084] Among them, a si has been normalized.

[0085] 3.2. Metrics of language area and semantic alignment development To measure how closely activations are related to specific language information, we define Linguistic Regions Development Scores (LRDS). Specifically, LRDS measures the difference between the average similarity of samples from the same language and samples from different languages, where all pairs of samples have different semantics. The number of samples for each language is equal. Lower LRDS indicates weaker language correlation of activation patterns.

[0086]

[0087] The Key Language Region Size (SKLR) is the sum of the key language region sizes for all languages. The Key Language Region Size (SKLR) is a metric used to evaluate the computational cost of aligning the input of a specific language to the semantic space, which we will further introduce in Section 3.3.

[0088] To measure how closely the activations are to the input semantics rather than the input language, we define Semantic Alignment Development Scores (SADS). Specifically, SADS measures the average similarity difference between samples with the same meaning and samples with different meanings, with all pairs of samples in different languages. A higher SADS indicates that the activation pattern is more related to the semantics of the input.

[0089]

[0090] 3.3. Key language area detection We consider neurons that activate in similar patterns across different samples of a language to be key neurons for that particular language. Therefore, we score these neurons to assess their contribution to the average similarity of samples for that particular language.

[0091] A specific language l The average similarity of samples in It can be expressed as:

[0092] in, n It’s language l The total number of samples in It's a neuron k (Activation vector k components) for samples s i The activation amount, K is the total number of neurons (i.e. the length of the activation vector).

[0093] We can see that neurons k The contribution to the mean cosine similarity can be quantified using the term in brackets. Therefore, we have neuron k For a language l Contribution score Defined as:

[0094] Since some neurons are always active, their contribution to the average similarity is always large. Therefore, we need to find neurons that contribute particularly large amounts to the average similarity for a particular language. To do this, we calculate the number of neurons k Standardized scores of contribution scores in different languages ​​(z-score, ):

[0095] in, μ k and σ k Neurons k Mean and standard deviation of contribution scores across all languages:

[0096] here, L is the total number of languages. Representing neurons k Language lIn practice, we set a threshold for the z-score, and neurons with z-scores above this threshold for a particular language are considered to be part of the critical language region for that language. For example, if the threshold is set to 2, then we select neurons whose activation scores are more than 2 standard deviations above their mean score.

[0097] 4. Experiment We first examine the similarity of activation patterns within a single model. To illustrate the linguistic and semantic similarity of inputs processed by LLMs, we plotted the activation similarity of neurons when processing inputs of the same language or the same semantics.

[0098] Next, we use the obtained similarity patterns to identify the key language regions for each language. We evaluate the performance of LLM by deactivating the key regions for each language separately, demonstrating the existence and usefulness of these regions.

[0099] We then explore the dynamic changes during training and scaling. We compute the Language Region Development Score (LRDS), Key Language Region Size (SKLR), and Semantic Alignment Development Score (SADS) at different training checkpoints to illustrate the development of language regions and semantic alignment. We perform the same analysis on models of different sizes to show their evolutionary trends.

[0100] 4.1 Experimental Setup Models: The experiments mainly use BLOOM. We tested various models in this series, including 560m, 1.1b, 1.7b, 3B, and 7.1b models, as well as the intermediate checkpoint of BLOOM-7b1. In addition, we also conducted supplementary experiments on the LLaMA-2 model family to examine the case where multilingual training data is limited.

[0101] Dataset and Language Selection: We use the Bible dataset (fully aligned parallel corpus) to activate LLM. Each sample consists of a verse, and we randomly selected 100 verses and selected translations in different languages ​​from the dataset. To evaluate multilingual perplexity, we used the XL-Sum dataset. The XL-Sum dataset contains high-quality articles from the BBC and covers 45 languages. Our experiments focus on 9 language subsets in BLOOM and XL-Sum: Arabic (ar), Simplified Chinese (zh-Hans), English (en), French (fr), Hindi (hi), Indonesian (id), Portuguese (pt), Spanish (es), and Vietnamese (vi).

[0102] Evaluation: We evaluated the complexity of each language model separately using XL-Sum. In addition, we designed a task to evaluate the cross-lingual reasoning ability of LLMs using the well-established EleutherAI-eval-harness framework. To this end, we used the XStoryCloze dataset, which consists of a short story typically consisting of four sentences and two alternative endings. One of the endings logically completes the story, while the other does not. The model is tasked with finding the more plausible ending. We prompted the story in various languages ​​and asked the model to choose between the two English ending options. Prompting the story and endings in different languages ​​helps us determine whether the model can effectively "understand" the story and map them into a common semantic space, rather than reasoning in only one language. The prompt format and evaluation details are shown in Appendix A.

[0103] Figure 4 This is a similarity map of all sample pairs generated by the BLOOM-7b1 model in this application. Each block of 100 samples uses the same language, and samples using the same language form different light blocks. Samples with the same semantics form light bands along the diagonal lines of these light blocks.

[0104] 4.2. Observation 1: Neuronal activations show similar patterns when processing different inputs of the same language or with the same semantics.

[0105] To illustrate the linguistic and semantic similarity of LLM when processing input, we take the BLOOM-7b1 model as an example and plot the similarity map of all sample pairs. Figure 4 As shown in the figure, each sample block consists of 100 samples of the same language, and sentences with the same semantics are placed in the same position in all sample blocks. In each language block, the similarity of neuronal activation is significantly higher, indicating that when processing input of the same language, the neuronal activation of LLMs presents a similar pattern.

[0106] In addition, we observed bright bands on the diagonal of the off-diagonal blocks. These bright bands represent sentences in different languages ​​but with the same semantics. Even languages ​​like Chinese and Arabic, which use different alphabets and do not share common tokens with other languages, have similar activation patterns when processing semantically identical sentences. This suggests that LLMs “understand” the semantics of the sentences, encoding the input information into a common semantic space shared by all languages, and are thus able to process information from different languages ​​in a similar manner.

[0107] Interestingly, we also observed that linguistically similar languages ​​such as French, Spanish, and Portuguese showed higher cross-lingual similarity compared to other language pairs. This may be because these languages ​​share more tokens and have similar grammatical syntax, causing LLMs to be processed in a similar way.

[0108]

[0109] Table 1: Percentage increase in complexity after deactivating neurons in the critical language regions for each language in the BLOOM-7b1 model. Each column corresponds to deactivating the critical region for a specific language. The second column shows the results of randomly deactivating 10% of neurons in the LLMs.

[0110] 4.3. Observation 2: Similarities in activation patterns help locate key language areas for a particular language.

[0111] To identify the critical language regions for each language, we calculated the z-score of each neuron in each language and selected neurons with z-scores greater than 2 (i.e., their activation score in a particular language was more than two standard deviations above the mean). These selected neurons made an unusual contribution to the specific language analyzed. We then deactivated the critical neurons for a specific language and measured the perplexity of the model across all languages. The results are shown in Table 1.

[0112] We found that removing the critical region for one language significantly increased the perplexity for that language, as shown in the diagonal block in the table. However, the perplexity for the other languages ​​remained almost the same, as shown outside the diagonal. This suggests that the neurons we found are essential for processing the specific language studied and that they constitute the critical language region for that language.

[0113] 4.4. Observation 3: As the training process progresses, the language critical area becomes smaller and smaller, and the activation pattern becomes more language-independent.

[0114] Now we can compare the development of language regions throughout the training process. Based on observation 2, we can find the key language region for each language, count the number of key neurons, and thus determine the size of the key language region (SKLR), which is the sum of the sizes of the key language regions for all languages. We can also use the similarity graph to calculate the language region development score (LRDS). By applying these metrics to BLOOM's training checkpoints, we can observe the development dynamics of the key language regions. Figure 5a , which is a schematic diagram of the changes in the language area along with the BLOOM-7b1 training steps in this application.

[0115] At the beginning of the training process, the size of the key language region is large, indicating a high number of key neurons. At the same time, the LRDS are also high, indicating that the activation pattern is highly language-specific. However, as training progresses, the size of the key language region decreases and the LRDS decreases. This suggests that the activation pattern is less related to a specific language and more focused on the semantics of the input. This shift occurs because the model is becoming more familiar with the language, so it doesn't need to spend as much effort to "understand" the sentence and project it into a common semantic space.

[0116] If we study the distribution of key neurons, we will find that key neurons are generally located in the first layer and the last few layers. Figure 6 As shown, it is a schematic diagram of the distribution of key neurons in different layers of BLOOM7b1 in this application. Among them, as the number of training steps increases, these key areas become more dense in the first few layers. The first few layers may be related to the encoding process from the source language to the common semantic space, while the latter layers correspond to the decoding process from the latent semantic representation to the target language. As training proceeds, the model can more effectively encode information from different languages ​​into the semantic space, and then use fewer neurons to decode this representation to the target language.

[0117] 4.5. Observation 4: As the training process progresses, the semantic alignment phenomenon becomes more obvious.

[0118] like Figure 5b As shown in the figure, it is a schematic diagram of the change of semantic alignment with the training steps of BLOOM7b1 in this application. In order to examine the semantic alignment phenomenon throughout the training process, we evaluated the semantic alignment development score (SADS) and cross-lingual zero-shot task performance of different checkpoints of BLOOM-7b1. The results are shown in Figure 5b As shown in the figure, as the training process progresses, SADS continues to increase and the model's cross-lingual reasoning ability continues to improve. This shows that the activations are increasingly related to the semantic rather than linguistic information of the input. Therefore, the model can gradually better integrate the input information into the common semantic space of different languages.

[0119] like Figure 7 As shown in FIG. 1 , the distribution of key neurons in different languages ​​on BLOOM-7b1 in this application is shown. Figure 7 As shown in Figure 1, by observing the size of the key language regions for each language, we can see that languages ​​with less similar linguistic features, such as Vietnamese, Arabic, and Chinese, have much larger key language regions than similar languages ​​because they belong to completely different language families and use different writing systems. This suggests that aligning these languages ​​in a common semantic space is generally more challenging because the model needs to spend more effort to align these languages ​​due to their different representations and linguistic features.

[0120] 4.6. Observation 5: As the model size increases, it can better align the language with the semantic space.

[0121] like Figure 8 Shown is a schematic diagram of the metrics of models of different sizes across the BLOOM family in this application. We evaluated the LRDS, SADS, and cross-lingual zero-shot task accuracy of BLOOM models of different sizes. We can clearly see that as the model size increases, the activation pattern of neurons becomes less and less related to the language and more focused on the semantics of the input. At the same time, the cross-lingual reasoning ability of the model also improves with the increase in scale. We conclude that as the model size increases, LLMs can better adjust inputs from different languages ​​to a common semantic space, thereby enhancing their reasoning process in this shared semantic latent space.

[0122] 4.7. Extension to LLaMA2 We perform the same experiments on LLaMA2. Since the intermediate checkpoints of LLaMA2 have not yet been released, we analyze different sizes of models, specifically 7b, 13b, and 70b models. We obtain similar results as the experiments conducted on BLOOM, namely the impact of the presence of language regions, semantic alignment, and the size of the enlarged language model. The results are shown in Appendix B.

[0123] 5. Related Work Multilingual Large Models. Large language models (LLMs) such as GPT-4, LLaMA, and OPT have revolutionized natural language processing by demonstrating the ability to understand and generate nuanced text. Multilingual LLMs such as BLOOM and XLM-R have further extended these capabilities by learning universal representations from text in multiple languages. These models typically use a transformer architecture to incorporate multilingual data during pre-training to improve cross-lingual alignment and performance. Various adaptation strategies have been proposed to adjust capabilities for different languages, including adaptation during pre-training, supervised fine-tuning, reinforcement learning from human feedback, downstream fine-tuning, just-in-time adaptation, and contrastive learning.

[0124] Neuroscience. In the field of neuroscience, a great deal of research has been conducted to understand how the human brain processes multiple languages. Studies have shown that polyglots (people who speak more than one language) show different brain activation patterns for different languages. These findings suggest that different languages ​​are stored and processed in different areas of the brain. Additionally, similar brain activation patterns may appear in different brain regions, suggesting that there is a common neural mechanism in multilingual processing.

[0125] Language and semantic alignment. Previous studies have shown that the representations generated by multilingual encoder models are to some extent language-independent. Based on this assumption, the LangBridge model is introduced to connect the multilingual encoder with monolingual LLMs, which effectively achieves good performance. It is also found that the efficacy of LangBridge stems from the language-independent nature of multilingual representations. It is further proposed to transfer English embeddings to virtual multilingual embeddings without semantic loss, thereby improving cross-lingual transfer capabilities. In addition, a core region corresponding to language consistency is discovered, and freezing the core language region in further pre-training can alleviate the problem of catastrophic forgetting.

[0126] 6. Conclusion This application explores the internal mechanisms of multilingual LLMs. We found that the activation patterns of neurons in LLMs are similar when processing the same language, which allows us to identify key neurons for specific languages. Inactivating these neurons significantly affects the performance of these languages. In addition, we found that the activation patterns of neurons are also similar when processing semantically identical sentences in different languages. This suggests that LLMs map these inputs to a common latent space. As training progresses, the key language region becomes smaller and the activation of neurons focuses more on semantics rather than language specific details. The key neurons are mainly located in the first and last layers, and the neurons in the first layer become denser as training progresses. In addition, larger models can better coordinate languages, thereby improving cross-lingual reasoning capabilities. Our results deeply reveal the structural evolution of multilingual LLMs during training and expansion, laying the foundation for improving their cross-lingual capabilities.

[0127] Appendix A: XStoryCloze prompt format We test the log likelihood of generating each option using a specific prompt. If the log likelihood of generating the correct ending is high, we infer that the model has correctly understood the story.

[0128] hint: [input_sentence_1,input_sentence_2,input_sentence_3,input_sentence_4,"The ending in English:"] choose: [Ending 1 in English, Ending 2 in English] We computed the log-likelihood of the model that produced the two choices after the prompt.

[0129] Example: 1. The mother told the children that it would be lunchtime in fifteen minutes. 2. Pero,entons,recibióuna llamada importante.(But then she got an important phone call.);Note Chinese translation: But then she got an important phone call 3. While she was talking, the dog dragged mud all over the kitchen. (While she was talking, the dog dragged mud all over the kitchen.) 4. Los The kids started to pester their mother, who was still on the phone. The ending in English: 1.The mother felt quite frustrated. 2.The children's behavior calmed the mother down. Correct Ending:1 Example Prompt:La madre ledijo a sus hijos que comerían en quinceminutos.Pero,entonces,recibióuna llamada importante.Mientras hablaba,el perrollenótoda la cocina de barro.Los English translation: Mom told the children that lunch was in fifteen minutes. But then she got an important phone call. As she spoke, the dog dragged mud all over the kitchen. The children began to pester their mother who was still on the phone. The ending in English: Choice 1(Target choice):The mother felt quite frustrated. Choice 2:The children's behavior calmed the mother down. The dataset supports evaluation for Arabic, Basque, Chinese, English, Hindi, Indonesian, Malay, Russian, Spanish, Swahili, and Telugu.

[0130] Appendix B: Results of LLaMA2 B.1 Similarity Graph like Figure 9a As shown, it is the similarity map of the LLaMA2-7b model in this application; Figure 9b As shown, it is the similarity map of the LLaMA2-13b model in this application; Figure 9b The figure shows the similarity map of the LLaMA2-70b model in this application. We can clearly see the light-colored blocks representing different languages ​​and the light-colored bands of semantically identical sentences. However, since LLaMA2 only used about 10% non-English data during training, they are not as prominent as the BLOOM model series.

[0131] B.2. Key language areas Table B.2 lists the percentage of complexity increase after deactivating neurons in key language regions for each language in the LLaMA2-7b model. As with the BLOOM model, we can see that there are some key language regions in LLaMA2-7b. When these regions are deactivated, the perplexity of the corresponding language increases significantly, while the perplexity of other languages ​​remains basically unchanged.

[0132] B.3 Comparison of models at different scales We evaluated the LLaMA2 results at different scales ( Fig.10 ). We can see that as the model size increases, the activation patterns become more language-agnostic and semantically focused, just like the BLOOM model family.

[0133]

[0134] Table 2: Percent increase in perplexity after deactivating neurons in the critical language regions for each language in the LLaMA2-7b model. Each column corresponds to deactivating the critical region for a specific language. The last column shows the results of randomly deactivating 17% of neurons in the LLM (slightly higher than the maximum percentage of critical neurons, vi is 16.5%).

[0135] like Fig.10 As shown, it is a schematic diagram of the metrics of different size models of the LLaMA2 series in this application. As the model size increases, the activation becomes more language-independent and semantically more focused.

[0136] like Fig.11 Shown is a schematic diagram of the distribution of LLaMA2-7b key neurons in different languages ​​in this application.

[0137] The computer device of the embodiment of the present application exists in various forms, including but not limited to: (1) Mobile communication devices: These devices are characterized by their mobile communication functions and their main purpose is to provide voice and data communications. These terminals include: smart phones (such as iPhone), multimedia phones, functional phones, and low-end phones.

[0138] (2) Ultra-mobile personal computer devices: These devices fall into the category of personal computers, have computing and processing capabilities, and generally also have mobile Internet access features. These terminals include: PDA, MID and UMPC devices, such as iPad.

[0139] (3) Portable entertainment devices: These devices can display and play multimedia content. They include audio and video players (such as iPods), handheld game consoles, e-books, as well as smart toys and portable car navigation devices.

[0140] (4) Server: A device that provides computing services. The server consists of a processor, hard disk, memory, system bus, etc. The server is similar to a general computer architecture, but because it needs to provide highly reliable services, it has higher requirements in terms of processing power, stability, reliability, security, scalability, and manageability.

[0141] (5) Other electronic devices with data interaction functions.

[0142] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0143] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a general hardware platform, and of course, by hardware. Based on this understanding, the above technical solution, in essence, or the part that contributes to the relevant technology, can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiment.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A multi-language large model optimization training method, comprising: Acquire a multilingual dataset of multiple languages ​​supported by the first initial multilingual large model; Inputting at least one language data in the multilingual data set into the first initial multilingual large model to determine a neuron set corresponding to an optimized training target; Presetting the neuron parameters corresponding to the neuron set according to the optimization training objective to obtain a second initial multilingual large model; The second initial multilingual large model is trained using a training sample set corresponding to the optimization training objective.

2. The method according to claim 1, characterized in that The optimization training target includes adding new language processing capabilities to the multilingual large model; the neuron set corresponding to the optimization training target includes the key neuron set corresponding to the multiple languages ​​supported by the first initial multilingual large model.

3. The method according to claim 2, characterized in that Presetting the neuron parameters corresponding to the neuron set according to the optimization training objective to obtain a second initial multilingual large model, including: Freeze corresponding upstream and downstream weight parameters of key neuron sets corresponding to multiple languages ​​supported by the first initial multilingual large model to obtain a second initial multilingual large model.

4. The method according to claim 1, characterized in that: The optimization training target includes enhancing the processing capability of the multilingual large model for the old languages ​​that are already supported; the neuron set corresponding to the optimization training target includes the key neuron set corresponding to the old languages ​​supported by the first initial multilingual large model.

5. The method according to claim 4, characterized in that Presetting the neuron parameters corresponding to the neuron set according to the optimization training objective to obtain a second initial multilingual large model, including: Freeze the corresponding upstream and downstream weight parameters of the key neuron set corresponding to the old language supported by the first initial multilingual large model to obtain a second initial multilingual large model.

6. The method according to claim 1, characterized in that The optimization training target includes the processing capability of the forgotten multilingual large model for the supported old languages; the neuron set corresponding to the optimization training target includes the key neuron set corresponding to the old languages ​​that need to be forgotten in the first initial multilingual large model.

7. The method according to claim 6, characterized in that Presetting the neuron parameters corresponding to the neuron set according to the optimization training objective to obtain a second initial multilingual large model, including: Reset corresponding upstream and downstream weight parameters of the key neuron set corresponding to the old language that needs to be forgotten in the first initial multilingual large model to zero to obtain a second initial multilingual large model.

8. A computer device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images