Method for constructing myanmar-chinese parallel corpus based on multi-step thinking large model

Abstract

This invention relates to a method for constructing a large-scale Burmese-Chinese parallel corpus based on multi-step thinking. The invention includes: translating existing Chinese-English parallel corpora using currently available translation models to obtain original English-Burmese and Chinese-Burmese parallel sentence pairs; calculating double-confidence intervals for semantic similarity and perplexity between aligned Burmese-Chinese sentence pairs based on publicly available high-quality Burmese-Chinese parallel corpora; using the selected double-confidence intervals to perform preliminary screening of Burmese-Chinese parallel sentence pairs, forming pre-processed pseudo-parallel sentence pairs; designing a multi-step thinking chain to guide the large-scale model to progressively optimize the pre-processed pseudo-parallel sentence pairs, thereby generating high-quality Burmese-Chinese parallel corpora for training the translation model, thus effectively improving the performance of Burmese-Chinese machine translation. This invention significantly enhances the ability of large language models to construct corpora in Burmese, a low-resource language, and provides an interpretable and transferable technical paradigm for corpus construction in other low-resource languages.

Description

[0001] Technical Language

[0002] This invention relates to a method for constructing a large-scale Myanmar-Chinese parallel corpus based on multi-step thinking, belonging to the field of natural language processing technology. Background Technology

[0003] Against the backdrop of accelerating globalization and the construction of the China-Myanmar Economic Corridor, the importance of China and Myanmar in economic and trade cooperation and cultural exchanges is becoming increasingly prominent. With the deepening of bilateral exchanges, the demand for Chinese-Burmese translation has exploded. However, Burmese is a typical low-resource language, and existing Burmese-Chinese parallel corpora suffer from severe problems such as resource scarcity and inconsistent quality, seriously restricting the quality and efficiency of machine translation. In practical translation scenarios, due to Burmese's unique grammatical structure and complex character encoding, existing translation systems often exhibit semantic deviations or word order confusion. Traditional Chinese-Burmese corpus construction mainly relies on manual collection, translation, and proofreading, a method that is inefficient and limited by human knowledge and effort, making it difficult to achieve large-scale, high-quality corpus production. Although general-purpose translation tools can provide basic assistance, they often fall short when dealing with specialized terminology and complex sentences.

[0004] In recent years, large-scale models such as the GPT series and Llama have made groundbreaking progress in natural language processing tasks. These models, trained on massive amounts of data, are able to grasp the general rules of language. However, directly applying these large models to Burmese-Chinese translation still faces challenges. Models are prone to factual illusions on low-resource languages, and the lack of effective verification mechanisms for single translation paths easily leads to error propagation. Against this backdrop, corpus construction methods utilizing multi-source data for cross-validation have gained increasing attention. By introducing English as an auxiliary source language to generate dual-path candidate translations, combining statistical features to construct a quality filter, and leveraging the reasoning capabilities of large-scale models for deep data optimization and error correction, the shortcomings of traditional methods can be effectively overcome. This method not only filters out corpora with high semantic consistency but also corrects for the grammatical characteristics of Burmese, significantly improving the quality and practicality of parallel corpora. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method for constructing a large-scale Myanmar-Chinese parallel corpus based on multi-step thinking. This method constructs a Myanmar-Chinese parallel corpus, selects corpora with high semantic consistency, and significantly improves the quality and practicality of the parallel corpus.

[0006] To achieve the above objectives, the technical solution adopted by this invention is as follows: a method for constructing a large-scale Burmese-Chinese parallel corpus based on multi-step thinking, characterized in that the method includes:

[0007] Step 1: Using a general machine translation model, translate the Chinese and English texts in the Chinese-English parallel corpus into Burmese respectively, and construct a candidate parallel dataset containing the original Chinese and dual-source Burmese candidate translations;

[0008] Step 2: Based on high-quality human-translated Myanmar-China parallel corpus, calculate the distribution features of semantic similarity and perplexity, construct a double confidence interval for semantic similarity and perplexity, and use the double confidence interval to strictly screen the candidate dataset generated in the previous step, retaining the pseudo-parallel sentence pairs that fall within the interval at the beginning of the process.

[0009] Step 3: Using a large language model based on multi-step thinking chain technology, the Burmese part of the selected pseudo-parallel sentence pairs is deeply optimized. Through the selection of the best or the fusion generation mode, the final high-quality Burmese-Chinese parallel corpus is output.

[0010] Step 4: After cleaning the generated corpus, use it to train the Burmese-Chinese machine translation model, and evaluate the model's translation performance from Burmese to Chinese.

[0011] Furthermore, the specific implementation of Step 1 includes:

[0012] Step 1.1: Collect and preprocess publicly available Chinese-English parallel corpora, and clean and standardize the data format;

[0013] Step 1.2: Deploy or invoke a general machine translation model to translate the Chinese sentences in the Chinese-English corpus as the target reference into the first Burmese candidate sentence, and translate the English sentences in the Chinese-English corpus into the second Burmese candidate sentence;

[0014] Step 1.3: Align and associate the original Chinese sentence with the two Burmese candidate sentences to construct a triplet dual-source candidate dataset containing the original Chinese sentence, the first Burmese candidate sentence, and the second Burmese candidate sentence.

[0015] Furthermore, the specific implementation of Step 2 includes:

[0016] Step 2.1: Select high-quality artificial Burmese-Chinese parallel corpora and calculate their semantic similarity and perplexity distributions respectively; based on the statistical mean and standard deviation features, set a lower threshold for semantic similarity and construct an effective value range from the lower threshold to the maximum value; set an upper threshold for perplexity and construct an effective value range from the minimum value to the upper threshold; combine the two effective value ranges to form a double confidence interval;

[0017] Step 2.2: Calculate the semantic similarity and perplexity of two Burmese candidate sentences and the original Chinese sentences in the candidate dataset. If both indicators fall within the double confidence interval, the sentence pair is retained as a pseudo-parallel sentence pair for preliminary processing; otherwise, it is discarded.

[0018] Furthermore, the specific implementation of Step 3 includes:

[0019] Step 3.1: Construct prompt words containing character settings and multi-step thought chain instructions, and input the pseudo-parallel sentence pairs retained in Step 2 into the large language model;

[0020] Step 3.2: Guide the model to perform stepwise optimization operations: First, analyze the grammatical structure and semantic accuracy of the two Burmese candidate sentences, focusing on identifying the subject-object-verb word order features of Burmese and possible semantic illusions;

[0021] Step 3.3: Output modes include: directly selecting the better translation from the two candidate sentences, or when both candidate sentences have advantages and disadvantages, extracting the correct segments from each and fixing grammatical errors, and merging them to generate a new high-quality Burmese translation.

[0022] Furthermore, the specific implementation of Step 4 includes:

[0023] Step 4.1: Merge the Burmese corpus optimized by the large model with the corresponding original Chinese sentences, perform deep data cleaning, and remove duplicate sentence pairs, sentence pairs with abnormal length ratios, special symbols and invalid character combinations according to the established rules to form a high-quality Burmese-Chinese synthetic parallel corpus;

[0024] Step 4.2: Use the Myanmar-Chinese synthetic parallel corpus generated by this method to train or fine-tune the translation model, and use the publicly available test set to verify the performance of the model. Quantitatively evaluate the improvement of the Myanmar-Chinese translation performance of the translation model by calculating BLEU and chrF values.

[0025] The present invention also provides a large-scale Burmese-Chinese parallel corpus construction system based on multi-step thinking, the system comprising: a module for executing the large-scale Burmese-Chinese parallel corpus construction method based on multi-step thinking.

[0026] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the method for constructing a large-scale Myanmar-Chinese parallel corpus based on multi-step thinking.

[0027] The present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method for constructing a large-scale Myanmar-Chinese parallel corpus based on multi-step thinking.

[0028] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the method for constructing a large-scale Myanmar-Chinese parallel corpus based on multi-step thinking.

[0029] The beneficial effects of this invention are:

[0030] 1. This invention generates candidate translations by introducing Chinese-English dual-source data back-translation and combines a dual-confidence interval filtering mechanism based on semantic similarity and perplexity. This effectively overcomes the semantic bias caused by single-source translation and the illusion problem common in low-resource languages. The intersection screening of the dual indicators greatly enhances the accuracy of the initial corpus screening and effectively improves the semantic consistency and quality lower limit of pseudo-parallel sentence pairs.

[0031] 2. By employing automated double-confidence interval screening and multi-step optimization of large models, this invention significantly reduces the reliance on scarce Burmese professional human translation and proofreading, reduces the high cost and time consumption of manually constructing corpora, and significantly improves the automation level and efficiency of low-resource language data production, providing a replicable paradigm for the large-scale production of minority language corpora.

[0032] 3. This invention utilizes the multi-step thinking chain reasoning technology of large language models to selectively screen or merge candidate data from multiple sources. This not only specifically corrects subject-object-verb word order errors in Burmese, but also integrates the expressive advantages of different source languages, enabling the generated parallel corpus to exhibit better fluency and authenticity when dealing with complex Burmese grammatical structures.

[0033] 4. This invention, through deep cleaning and multi-step iterative optimization of the synthetic corpus, can construct a large-scale, high-quality Burmese-Chinese synthetic parallel translation corpus, effectively solving the problem of scarce Burmese data. Experiments have shown that this corpus can significantly improve the BLEU and chrF values ​​of the Burmese-Chinese machine translation model, providing a solid data foundation and technical support for breakthroughs in the performance of low-resource language translation. Attached Figure Description

[0034] Figure 1 This is a flowchart from the present invention. Detailed Implementation

[0035] Example 1: As Figure 1 As shown, a method for constructing a large-scale Myanmar-Chinese parallel corpus based on multi-step thinking includes the following steps:

[0036] Step 1: Using a general machine translation model, translate the Chinese and English texts in the Chinese-English parallel corpus into Burmese, respectively, and construct a candidate parallel dataset containing the original Chinese and dual-source Burmese candidate translations. The specific steps are as follows:

[0037] Step 1.1: Collect and preprocess 150,000 pairs of Chinese-English parallel corpora from the publicly available ParaCrawl Chinese-English parallel dataset. After cleaning and formatting the data, an initial corpus of nearly 130,000 corpora is obtained. This embodiment chooses the ParaCrawl dataset instead of the more commonly used WMT dataset because ParaCrawl is a web-based dataset that covers multiple areas of daily life. Although the noise ratio is slightly higher, its data characteristics are closer to the real-world application scenarios of low-resource languages, which can effectively verify the noise resistance and robustness of the method of this invention.

[0038] Step 1.2: Deploy or invoke the NLLB-200-600M multilingual machine translation model to translate the Chinese sentences in the Chinese-English corpus into first Burmese candidate sentences, and to translate the English sentences in the Chinese-English corpus into second Burmese candidate sentences.

[0039] Step 1.3: Align and associate the original Chinese sentence with the two Burmese candidate sentences to construct a triplet dual-source candidate dataset containing the original Chinese sentence, the first Burmese candidate sentence, and the second Burmese candidate sentence.

[0040] Step 2: Based on high-quality manually translated Myanmar-China parallel corpus, calculate the distribution features of semantic similarity and perplexity, construct double confidence intervals for semantic similarity and perplexity, and use these intervals to rigorously filter the candidate dataset generated in the previous step, retaining the pseudo-parallel sentence pairs that fall within both intervals after preliminary processing. The specific steps are as follows:

[0041] Step 2.1: Select 10,000 high-quality artificially generated Burmese-Chinese parallel corpora from the ALT dataset, calculate their semantic similarity and perplexity indices, and determine the effective value range of the high-quality corpora based on the statistical mean and standard deviation. Set a lower threshold for semantic similarity, constructing a retention interval from the lower threshold to the maximum value; set an upper threshold for perplexity, constructing a retention interval from the minimum value to the upper threshold. Combine these two retention intervals to determine the final double-confidence interval.

[0042] Step 2.2: Calculate the semantic similarity and perplexity of two Burmese candidate sentences and the original Chinese sentences in the candidate dataset. Only retain sentences when both indicators fall within the double confidence interval; otherwise, discard them to obtain the preliminary processed pseudo-parallel sentence pairs.

[0043] Step 3: Using the GPT-4o-mini large language model based on multi-step thinking chain technology, the Burmese portion of the selected pseudo-parallel sentence pairs is deeply optimized. Through optimal selection or fusion generation, the final high-quality Burmese-Chinese parallel corpus is output. The specific steps are as follows:

[0044] Step 3.1: Construct prompt words that include character settings and multi-step thought chain instructions, and input the pseudo-parallel sentence pairs retained in Step 2 into the large language model.

[0045] Step 3.2: Guide the GPT-4o-mini large language model to perform stepwise optimization. The model first analyzes the grammatical structure and semantic accuracy of the two Burmese candidate sentences, focusing on identifying and correcting segments that do not conform to the Burmese subject-object-verb word order, as well as possible semantic illusions or logical errors.

[0046] Step 3.3: Output modes include: directly selecting the better translation from two candidate sentences, or, when both candidate sentences have advantages and disadvantages, extracting the correct segments from each and correcting grammatical errors, then merging them to generate a new high-quality Burmese translation. The thought process template for the optimal selection mode is shown in Table 1, and the thought process template for the fusion generation mode is shown in Table 2.

[0047] Table 1. Mental Chain Template for Guiding Optimal Selection in Large Models

[0048] Table 2. Mental Chain Template for Guiding Optimal Selection in Large Models

[0049] Step 4: After cleaning the generated corpus, use it to train the Burmese-Chinese machine translation model, and evaluate the model's Burmese-to-Chinese translation performance. The specific steps are as follows:

[0050] Step 4.1: Merge the optimized Burmese corpus with the corresponding original Chinese sentences, perform deep data cleaning, and remove duplicate sentence pairs, sentence pairs with abnormal length ratios, special symbols, invalid character combinations, low similarity sentence pairs, and redundant labels according to established rules to form a high-quality Burmese-Chinese synthetic parallel corpus.

[0051] Step 4.2: The Myanmar-Chinese synthetic parallel corpus generated by this method is used to train the NLLB-200-600M base model to obtain an optimized Myanmar-Chinese machine translation model. The performance of the model is validated using the publicly available test set flores200. The performance improvement of the translation model is quantitatively evaluated by calculating the BLEU and chrF values ​​of the generated translation and the reference translation.

[0052] This invention uses the Bleu value and chrF value, which are commonly used in the field of error correction, as evaluation indicators for the model. The specific calculation method of the Bleu value is shown in formula (1):

[0053] (1)

[0054] in,

[0055] : Improved n-gram precision. Indicates the proportion of n-grams in the candidate translation that appear in the reference translation.

[0056] N: The maximum length of the n-gram, usually N=4.

[0057] Weighting coefficients, typically using uniform weighting, i.e. =1 / N.

[0058] BP: Shortness Penalty Coefficient, used to penalize generated translations shorter than the reference translation. The formula for calculating the Shortness Penalty Coefficient (BP) is as follows:

[0059] (2)

[0060] in:

[0061] c: Effective length of candidate sentences generated by machine translation

[0062] r: The effective length of the reference translation.

[0063] The specific method for calculating the chrF value is shown in formula (3):

[0064] (3)

[0065] Character-level n-gram precision. This refers to the proportion of n-grams in the candidate translation that match the reference translation out of the total number of n-grams in the candidate translation.

[0066] Character-level n-gram recall. This is the proportion of n-grams in the candidate translation that match the reference translation out of the total number of n-grams in the reference translation.

[0067] Weighting parameters are used to adjust the importance of precision and recall.

[0068] The present invention also provides a large-scale Burmese-Chinese parallel corpus construction system based on multi-step thinking, the system comprising: a module for executing the large-scale Burmese-Chinese parallel corpus construction method based on multi-step thinking.

[0069] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the method for constructing a large-scale Myanmar-Chinese parallel corpus based on multi-step thinking.

[0070] The present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method for constructing a large-scale Myanmar-Chinese parallel corpus based on multi-step thinking.

[0071] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the method for constructing a large-scale Myanmar-Chinese parallel corpus based on multi-step thinking.

[0072] This invention inputs the constructed parallel corpus into the translation model for experimentation, and also sets up multiple sets of comparative experiments to verify the effectiveness of the proposed optimization strategies at different stages (back-translation, statistical filtering, large model selection and fusion) in improving the performance of low-resource Myanmar-Chinese machine translation. Table 3 shows the comparison results of the impact of different corpus construction strategies on the performance of the final translation model.

[0073] The table shows:

[0074] The original NLLB model refers to inference using only the existing NLLB-200-600M basic model. While this basic model possesses broad multilingual translation capabilities, it suffers from limitations in handling complex Burmese grammatical structures such as SOV word order and polysemous words when faced with specific low-resource Burmese-Chinese translation tasks due to the lack of domain-specific parallel corpora for supervision. Consequently, its BLEU score is only 17.5, and its chrF score is only 16.5, both the lowest in the entire table. Chinese single-source back-translation and English single-source back-translation refer to the results of back-translating Chinese and English monolingual corpora into Burmese using a general translation model, and then training the model with the resulting 129,979 sentences as pseudo-parallel corpora. Experiments show that introducing large-scale back-translation data significantly improves model performance, with the BLEU score increasing to between 19.9 and 20.7, indicating that the expansion of the data scale compensates for the shortcomings of the basic model to some extent. The Chinese back-translation + statistical filtering and English back-translation + statistical filtering refer to the results of cleaning the original back-translated data by introducing the confidence interval filter in Step 2 of this invention before model training. Although the training data size was reduced from 129,979 to 84,055, the model performance remained stable or even slightly improved. For example, the chrF value of English back-translation + statistical filtering reached 20.1. This verifies that the statistical filtering mechanism of this invention can effectively remove noisy data in the back-translation process, achieving a training effect of reducing quantity and improving quality.

[0075] The dual-source selection and fusion generation of the large-scale model represent the final performance improvements achieved by further optimizing the large-scale model based on high-quality data that has undergone statistical filtering. The former is the corpus constructed after selecting candidate sentences from two sources using a large-scale model combined with thought chains; while the latter is the corpus generated by fusion generation using the large-scale model. Data shows that the model fine-tuned based on the fusion-generated corpus of the large-scale model achieved the best performance, with a BLEU score of 21.3 and a chrF score of 20.6. Compared with the original NLLB model, the BLEU score improved significantly by 3.8 points; compared with Chinese back-translation + statistical selection that only performs simple filtering, the BLEU score also showed a significant gain of 0.6 points. This strongly demonstrates that the fusion generation strategy proposed in this invention is not merely a simple data selection, but rather creates a high-quality corpus that far exceeds the generation capabilities of the basic model by guiding the large-scale model to absorb the strengths of multi-source translations, correct grammatical errors, and reorganize and generate new data. Ultimately, this high-quality synthetic corpus successfully adapted a general-purpose lightweight basic model into a high-performance dedicated translation model, verifying the enormous application potential of this method in low-resource language scenarios.

[0076] Table 3. Results of the comparative experiment on Burmese-Chinese translation.

[0077] Finally, through the synergistic effect of candidate data generated by bi-Chinese-English back-translation and deep optimization of the multi-step thinking chain of the large model, this invention significantly improves the accuracy and fluency of the machine translation model in low-resource Burmese-Chinese translation tasks. This achievement provides a replicable technical paradigm for translating low-resource languages ​​with scarce corpora and complex grammar, and verifies the feasibility of using dual-confidence interval statistical constraints and large-model inference to collaboratively solve the problems of data scarcity and uncontrollable quality of pseudo-parallel corpora in minority languages. The dual mechanism of dual-confidence interval filtering and thinking chain fusion not only enhances the accurate correction of semantic illusions and grammatical errors in Burmese, but also opens up new technical avenues for constructing high-quality synthetic corpora for other homogeneous low-resource languages.

[0078] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for constructing a large-scale Burmese-Chinese parallel corpus based on multi-step thinking, characterized in that: The method includes: Step 1: Using a general machine translation model, translate the Chinese and English texts in the Chinese-English parallel corpus into Burmese respectively, and construct a candidate parallel dataset containing the original Chinese and dual-source Burmese candidate translations; Step 2: Based on the manually translated Myanmar-China parallel corpus, calculate the distribution features of semantic similarity and perplexity, construct a double confidence interval for semantic similarity and perplexity, and use the double confidence interval to strictly screen the candidate dataset generated in the previous step, retaining the pseudo-parallel sentence pairs that fall within the interval in the preliminary processing. Step 3: Using a large language model based on multi-step thinking chain technology, optimize the Burmese part of the selected pseudo-parallel sentence pairs, and output the final Burmese-Chinese parallel corpus by selecting the best or merging generation mode.

2. The method for constructing a large-scale Myanmar-Chinese parallel corpus based on multi-step thinking according to claim 1, characterized in that... The specific implementation of Step 1 includes: Step 1.1: Collect and preprocess publicly available Chinese-English parallel corpora, and clean and standardize the data format; Step 1.2: Deploy or invoke a general machine translation model to translate the Chinese sentences in the Chinese-English corpus as the target reference into the first Burmese candidate sentence, and translate the English sentences in the Chinese-English corpus into the second Burmese candidate sentence; Step 1.3: Align and associate the original Chinese sentence with the two Burmese candidate sentences to construct a triplet dual-source candidate dataset containing the original Chinese sentence, the first Burmese candidate sentence, and the second Burmese candidate sentence.

3. The method according to claim 1, characterized in that... The specific implementation of Step 2 includes: Step 2.1: Select artificial Burmese-Chinese parallel corpora and calculate their semantic similarity and perplexity distributions respectively; based on the statistical mean and standard deviation features, set a lower threshold for semantic similarity and construct an effective value range from the lower threshold to the maximum value; set an upper threshold for perplexity and construct an effective value range from the minimum value to the upper threshold; combine the two effective value ranges to form a double confidence interval; Step 2.2: Calculate the semantic similarity and perplexity of two Burmese candidate sentences and the original Chinese sentences in the candidate dataset. If both indicators fall within the double confidence interval, the sentence pair is retained as a pseudo-parallel sentence pair for preliminary processing; otherwise, it is discarded.

4. The method according to claim 1, characterized in that... The specific implementation of Step 3 includes: Step 3.1: Construct prompt words containing character settings and multi-step thought chain instructions, and input the pseudo-parallel sentence pairs retained in Step 2 into the large language model; Step 3.2: Guide the model to perform stepwise optimization operations: First, analyze the grammatical structure and semantic accuracy of the two Burmese candidate sentences, focusing on identifying the subject-object-verb word order features of Burmese and possible semantic illusions; Step 3.3: Output modes include: directly selecting the better translation from the two candidate sentences, or when both candidate sentences have advantages and disadvantages, extracting the correct segments from each and fixing grammatical errors, and merging them to generate a new high-quality Burmese translation.

5. The method for constructing a large-scale Burmese-Chinese parallel corpus based on multi-step thinking according to claim 1, characterized in that, Also includes: Step 4: After cleaning the generated corpus, use it to train the Burmese-Chinese machine translation model and evaluate the model's Burmese-to-Chinese translation performance; Step 4 specifically includes: Step 4.1: Merge the Burmese corpus optimized by the large model with the corresponding original Chinese sentences, perform deep data cleaning, and remove duplicate sentence pairs, sentence pairs with abnormal length ratios, special symbols and invalid character combinations according to the established rules to form a high-quality Burmese-Chinese synthetic parallel corpus; Step 4.2: Use the Myanmar-Chinese synthetic parallel corpus generated by this method to train or fine-tune the translation model, and use the publicly available test set to verify the performance of the model. Quantitatively evaluate the improvement of the Myanmar-Chinese translation performance of the translation model by calculating BLEU and chrF values.

6. A large-scale model-based system for constructing Myanmar-China parallel corpora based on multi-step thinking, characterized in that: The system includes a module for performing the method for constructing a large-scale Myanmar-Chinese parallel corpus based on multi-step thinking as described in any one of claims 1 to 5.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the program, it implements the method for constructing a large-scale Myanmar-Chinese parallel corpus based on multi-step thinking as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for constructing a large-scale Myanmar-Chinese parallel corpus based on multi-step thinking as described in any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for constructing a large-scale Myanmar-Chinese parallel corpus based on multi-step thinking as described in any one of claims 1 to 5.

Images