A method for multi-language machine translation by multi-stage collaboration of generative multi-agents

Abstract

The present application relates to a kind of generation type multi-agent's multilingual machine translation method of multi-stage cooperation, belong to natural language processing field.It is specific to say, first, the traditional multilingual machine translation model is used to carry out beam decoding, forms N-best translation hypothesis list.Then, N-best hypothesis list and the translation of large language model are connected through multi-agent, complete multi-stage refinement, revision, selection, reach the accuracy and reliability of translation are improved.The present application combines the advantages of traditional multilingual translation model and large model through multi-agent architecture, by fully exploiting the potential of both, carry out extensive experiment on each language pair, and the result proves that the translation effect of the present application on each language pair reaches leading level, and has good universality.

Description

[0001] Technical language

[0002] The application relates to a multi-stage collaborative multi-language machine translation method of a generative multi-agent, and belongs to the technical field of natural language processing. BACKGROUND

[0003] In recent years, large language models (LLM), such as the GPT series, the LLaMA series, and the GLM series, have made significant progress in the field of natural language processing (NLP) and have become a research hotspot. Through self-supervised learning of massive text data, these models have shown amazing language understanding and generation capabilities, and have also shown great potential to drive the innovation of machine translation technology. Compared with traditional neural machine translation, the results of the general machine translation task in WMT23 show that LLM-based machine translation systems have achieved leading positions in most translation subtasks. These results show that LLM-based machine translation has become a new paradigm in the field of machine translation. However, it is worth noting that although the best-performing translation model in WMT23 is based on GPT-4, simple LLM-based direct translation systems still have many problems. Compared with traditional neural machine translation, LLM-based translation systems rely heavily on the ability of the model itself to output once, so their output stability is poor in machine translation tasks involving low resources. At the same time, it also naturally has some inherent defects of the large model itself, such as lack of interpretability and reproducibility, and hallucination problems. Therefore, how to consider combining the advantages of both to generate a high-quality translation is a noteworthy problem.

[0004] The application proposes a multi-stage collaborative multi-language machine translation method of a generative multi-agent, which uses multiple generative agents to deploy and cooperate with each other in multiple stages to complete the multi-language machine translation task. At the same time, it fully considers and maximizes the use of traditional machine translation models and results, effectively constrains the new translation paradigm of LLM-based machine translation systems, and provides translation information gain. In addition, through the multi-agent structure, this framework expands this influence, connects the traditional paradigm and the new paradigm in machine translation, and provides a balance point between the two. And on the basis of the two, facing the multi-language machine translation task, the model performance has reached a new height. SUMMARY

[0005] The application proposes a multi-stage collaborative multi-language machine translation method of a generative multi-agent to solve the problems of translation quality in low-resource traditional multi-language machine translation and the potential of large language models in translation tasks not being fully tapped. The application has achieved good experimental results on related data sets.

[0006] The technical solution of the present application is: a multi-stage collaborative multi-language machine translation method of generative multi-agent, the specific steps of the method are as follows:

[0007] Step1, select the traditional multi-language machine translation model mRASP2 as one of the basic model architectures, which is suitable for the needs of multi-language machine translation tasks, after selection, the mRASP2 model is tested and beam decoding is generated.

[0008] Step2, select a suitable large model, such as Llama3.1-8B-Instruct, GPT-4omini, as one of the basic model architectures, which needs to be suitable for the needs of multi-language machine translation tasks, after selection, the large model is started based on GMATranslate multi-agent expert.

[0009] Step3, use the multi-language translation expert under the GMATranslate framework to complete the first preliminary translation for the target to be translated;

[0010] Step4, use the information expert under the GMATranslate framework to summarize the information based on the results of mRASP2 beam decoding, and get the second preliminary translation;

[0011] Step5, use the review expert and refining expert under the GMATranslate framework to review and refine the preliminary translation 1 and the preliminary translation 2;

[0012] Step6, use the selection expert under the GMATranslate framework to select the final translation that best matches the content, emotion and meaning of the source text.

[0013] As a further scheme of the present application, the specific steps of Step1 are as follows:

[0014] Step1.1, obtain the open source and pre-trained multi-language translation model mRASP2, which has a twelve-layer encoding-decoding framework;

[0015] Step1.2, for the text to be translated S, first perform Tokenization processing, and then perform binary processing;

[0016] Step1.3, input the processed binary data into mRASP2, and generate an N-best translation hypothesis list about S through beam decoding

[0017] As a further scheme of the present application, the specific steps of Step2 are as follows:

[0018] Step 2.1: Obtain the open-source Llama 3.1-8B-Instruct as the base model for launching the multi-agent framework.

[0019] Step 2.2: Using the vllm framework, start Llama3.1-8B-Instruct locally and connect it to the various agents in the GMATranslate framework.

[0020] As a further aspect of the present invention, the specific steps of Step 3 are as follows:

[0021] Step 3.1: The multilingual translation expert in the GMATranslate framework receives the text to be translated S and the target language code as input to complete the first preliminary translation T1.

[0022] Step 3.2: Pass the first preliminary translation T1 as output information to the review experts.

[0023] As a further aspect of the present invention, the specific steps of Step 4 are as follows:

[0024] Step 4.1: Generate mRASP2 Using the text S to be translated as input, the information expert considers multiple dimensions of S, including content, sentiment, and style, to... The provided translation hypotheses are matched and summarized in multiple dimensions to obtain the second preliminary translation T2.

[0025] Step 4.2: Pass the output information, the second preliminary translation T2, to the review experts as input.

[0026] As a further aspect of the present invention, the specific steps of Step 5 are as follows:

[0027] Step 5.1: The review experts accept the first preliminary translation T1 and the second preliminary translation T2, as well as the text to be translated S, as input. Then, they conduct multi-dimensional MQM review on T1 and T2 to obtain the review results R1 and R2 as output.

[0028] Step 5.2, Refinement Expert: Accepts two sets of information (T1, R1) and (T2, R2) as input, modifies T1 and T2 according to the cases of R1 and R2, and obtains two refined results Ref1 and Ref2 as output.

[0029] As a further aspect of the present invention, the specific steps of Step 6 are as follows:

[0030] Step6.1, select experts to accept Ref1, Ref2 and S as input, then select the most suitable result Tsec as output by comparing the closeness of Ref1, Ref2 and S through secondary multi-dimensional comparison.

[0031] The beneficial effects of the present application are:

[0032] 1. The present application realizes the bridging of new and old paradigms in the field of multilingual machine translation, fully utilizes the advantages of both, and improves the translation performance;

[0033] 2. The present application has carried out a large number of tests on high-resource and low-resource corpora in multilingual, which greatly improves the low-quality translation of the basic model;

[0034] 3. The present application proves the effectiveness and universality of the present application through tests on different base models and different languages;

[0035] 4. The method proposed in the present application greatly improves the accuracy of translation and the closeness of semantics compared with the original model in extensive multilingual tests. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 is a schematic diagram of the model capabilities of the present application;

[0037] Figure 2 is a translation score diagram of each model of the present application;

[0038] Figure 3 is a flowchart in the present application. DETAILED DESCRIPTION

[0039] Embodiment 1: As shown, the method for generating multi-agent multi-stage cooperation to complete multilingual machine translation, the specific steps of the method are as follows: Figures 1-3

[0040] Step1, select the traditional multilingual machine translation model mRASP2 as the basic model architecture, test the effect of the mRASP2 model and beam decode to generate a translation hypothesis list.

[0041] ​Step 1.1, download mRASP2 model, and its pre-processing file package mRASP-dataprocess, then set the mRASP2 model parameters. The embedding size and FFN dimension are set to 1024, dropout = 0.1, learning rate is 3e-4, polynomial decay schedule and 10000 warm-up steps. For optimization, use the Adam optimizer, where β1 = 1e-6 and β2 = 0.98. For multi-language vocabulary, follow the shared BPE vocabulary, which includes 59 languages. The vocabulary contains 64808 tokens. After adding 59 language tokens, the total size of the vocabulary is 64867.

[0042] Step 1.2, download TED2020 dataset in OPUS database, which covers all six languages tested: English (EN), French (FR), German (DE), Chinese (ZH), Vietnamese (VI) and Malay (MS).

[0043] Step 1.3, Tokenization and binarization of test data.

[0044] Step 1.4, beam decoding generation using fairseq-generate on binarized data, then data cleaning on decoding results to get Translation hypothesis list.

[0045] Step 2, download Llama3.1-8B-Instruct to local, download vllm to local, use vllm to start Llama3.1-8B-Instruct, start GMATranslate framework.

[0046] Step 2.1, download Llama3.1-8B-Instruct on Huggingface to local, and download OPENAI package;

[0047] Step 2.2, download vllm to local, then start Llama3.1-8B-Instruct through vllm framework, port mapping to localhost:8080;

[0048] Step 2.3, download Dify to local, and start the previously constructed GMATranslate framework in Dify, load the multi-agent in GMATranslate as Llama3.1-8B-Instruct through configuring Openai API.

[0049] Step3、Using the multilingual translation expert under the GMATranslate framework, the first preliminary translation is completed for the target to be translated;

[0050] Step3.1、The original text S, the self-information I of the multilingual expert and the target language code Ti are input to the multilingual expert;

[0051] Step3.2、S, I and Ti are passed to the multilingual review expert, and then the multilingual expert outputs the first preliminary translation T1, as follows:

[0052] T1=DirectTranslate(S,T i ,I1) (1)

[0053] Step4、Using the information expert under the GMATranslate framework, the information summary is performed for the result of the beam decoding of mRASP2, and the information expert is used to summarize to obtain the second preliminary translation T2.

[0054] Step4.1、The generated by mRASP2 and the text to be translated S and the self-information I2 of the expert are input to the information expert.

[0055] Step4.2、According to the content, emotion, style and other multi-dimensional factors of S, the information expert performs multi-dimensional matching summary on the translation hypotheses provided by to obtain the preliminary translation T2, as follows:

[0056]

[0057] Step5、Revision and refinement are performed on T1 and T2.

[0058] Step5.1、T1, T2, S and I3 are input to the review expert to form two triples (T1, S, I3) and (T2, S, I3). Each triple is reviewed once, as follows:

[0059] R1=Revision(T1,S,I3) (3)

[0060] R2=Revision(T2,S,I3) (4)

[0061] R=Concat(R1,R2) (5)

[0062] Step5.2、T1, T2, R and I4 are input to the refinement expert to form a quadruple (T1, T2, R, I4). The quadruple is refined, as follows:

[0063] Ref = Refine(T1, T2, R, I4) (6)

[0064] Step 6: Select experts to accept Ref1, Ref2, and S as input, and then compare the similarity of Ref1, Ref2, and S in a second multi-dimensional manner to select the most suitable result Tsec as output. The specific formula is as follows:

[0065] Tsec = Select(Ref, S, I5) (7)

[0066] To illustrate the effectiveness of this invention, three evaluation metrics—COMET, COMETKiwi, and BLUERT—were used. Six languages ​​(English, French, German, Chinese, Malay, and Vietnamese) were tested, with English as the pivot language. A total of ten translation directions were generated, including: English to Chinese (EN-ZH), English to French (EN-FR), English to German (EN-DE), English to Vietnamese (EN-VI), English to Malay (EN-MS), Chinese to English (ZH-EN), French to English (FR-EN), German to English (DE-EN), Vietnamese to English (VI-EN), and Malay to English (MS-EN). Of these ten directions, four are low-resource translation directions, while the remaining six are high-resource translation directions. This arrangement helps to comprehensively evaluate the performance of this invention under different resource conditions, especially for low-resource languages ​​such as Vietnamese and Malay. The final experimental results are shown in Table 1 below.

[0067] Table 1 Comparison of translation performance of other methods and the method of the present invention in various languages.

[0068]

[0069]

[0070] It can be clearly observed that across ten translation directions, the GMATranslate framework of this invention has achieved significant growth compared to the traditional strong baseline model mRASP2 or the new translation paradigm based on large language models, regardless of whether the language pairs are high-resource or low-resource. In particular, in the five translation directions with English as the source language, it achieved an average growth of 2.09 points under the COMETKiwi evaluation metric.

[0071] Meanwhile, in order to gain a more intuitive understanding of the capabilities of each model, Figure 1 The results of each model in ten translation directions were visualized.

[0072] To verify the universality of the present application, the present application is tested in ten directions for different base large models, and the self-gain of the present application in different base models is shown. It is worth noting that the alpha version of the present application is based on the Llama3.1-8b-Instruct model, and the beta version is based on the GPT-4omini model. The results are shown in Table 2.

[0073] Table 2 Self-gain of GMATranslate alpha and GMATranslate beta in ten directions

[0074]

[0075]

[0076] To verify the effectiveness of the method modules of the present application, the relevant ablation experiments are carried out. Since the adjudication expert does not directly output the translation answer, but works together with the refining expert to optimize the translation result, the two are placed in the same module for consideration. The selection expert is only responsible for the selection process and does not directly affect the translation result, so it is taken as a basic module. The results are shown in Table 3.

[0077] Table 3 Ablation experiment of the modules designed by the present application

[0078]

[0079] The experimental data clearly shows that each expert has made contributions, especially the adjudication expert and the refining expert have played a significant role in improving the translation quality.

[0080] To prove the point that the present application has a significant advantage in the translation direction of low-resource languages, Figure 2 The scores of the one hundred test sentences in the EN-MS translation pair are two-dimensionally visualized to more intuitively understand the specific changes in the translation process of EN-MS, the translation and model performance improvement of the five models, and the most intuitive contribution of the present application.

[0081] As Figure 2As shown, as a traditional multilingual strong baseline model, there are many light color blocks in the translation score graph of mRASP2, which represents that there are many low-quality translations. In contrast, although Llama3.1-8B-Instruct model also has many low-quality translations compared with the more powerful GPT-4omini, its overall performance is still better than that of mRASP2 model. Under the GMATranslate framework of the present application, the number of such light color blocks is obviously improved whether it is GMATranslateα or GMATranslateβ. This further proves the universality and effectiveness of the GMA-Translate framework for improving low-quality translations.

[0082] The specific embodiments of the present application are described in detail above in conjunction with the accompanying drawings, but the present application is not limited to the above-described embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the spirit of the present application.

Claims

1. A method for multi-stage collaborative machine translation of multi-language by generative multi-agent, characterized in that: The specific steps of the method are as follows: Step 1, select the traditional multilingual machine translation model mRASP2 as one of the basic model architectures, after selection, test the effect of the mRASP2 model and beam decoding generation; Step 2, select a large model, after selection, start the GMATranslate multi-agent expert based on the large model; Step 3, use the multilingual translation expert under the GMATranslate framework to complete the first preliminary translation for the target to be translated; Step 4, use the information expert under the GMATranslate framework to summarize the information of the mRASP2 beam decoding result to obtain the second preliminary translation; Step 5, use the review expert and the refining expert under the GMATranslate framework to review and refine the first preliminary translation and the second preliminary translation; Step 6, use the selection expert under the GMATranslate framework to select the final translation that best matches the content, emotion and meaning of the source text.

2. The method of claim 1, wherein the method further comprises: The specific steps of Step 1 are as follows: Step 1.1, obtain the open source and pre-trained multilingual translation model mRASP2, which has a twelve-layer encoding-decoding framework; Step 1.2, for the text to be translated S, first perform Tokenization processing, and then perform binary processing; Step 1.3, input the processed binarized data as input of mRASP2 and generate N-best translation hypothesis list about S through beam decoding 3. The method of claim 1, wherein the method further comprises: The large model selected in Step 2 is Llama3.1-8B-Instruct, GPT-4omini.

4. The method of claim 1, wherein the method further comprises: The specific steps of Step 2 include the following: Step 2.1, obtain the open source Llama3.1-8B-Instruct as the base model for starting the multi-agent framework; Step 2.2, through the vllm framework, start Llama3.1-8B-Instruct locally and connect it to each agent in the GMATranslate framework.

5. The method of claim 1, wherein the method further comprises: The specific steps of Step 3 include the following: Step 3.1, the multilingual translation expert in the GMATranslate framework receives the information of the text to be translated S and the target language code as input, and completes the first preliminary translation T1; Step 3.2, the first preliminary translation T1 is output information and is passed to the review expert.

6. The method of claim 1, wherein the method further comprises: The specific steps of Step 4 include the following: Step 4.1: Generate mRASP2 Using the text S to be translated as input, the information expert analyzes and translates it based on multiple dimensions of S, including content, sentiment, and style. The provided translation hypotheses are matched and summarized in multiple dimensions to obtain the second preliminary translation T2. Step 4.2, the output information second preliminary translation T2 is passed to the review expert as input.

7. The method of claim 1, wherein the method further comprises: The specific steps of Step 5 include the following: Step 5.1, the review expert accepts the first preliminary translation T1 and the second preliminary translation T2 as well as the text to be translated S as input, then performs multi-dimensional MQM review on T1 and T2 to obtain the review results R1 and R2 as output; Step 5.2, the refining expert accepts two groups of information (T1, R1) and (T2, R2) as input, modifies T1 and T2 according to the situation of R1 and R2, and obtains two refining results Ref1 and Ref2 as output.

8. The method of claim 7, wherein the method further comprises: The specific steps of Step 6 are as follows: Step 6.1, select the expert to accept Ref1, Ref2 and S as input, and then through the secondary multi-dimensional comparison of the closeness of Ref1, Ref2 and S, select the most suitable result Tsec as output.

Images