Machine translation originated from research in Natural Language Processing (NLP), which is an intersection of linguistics and artificial intelligence, and can be seen as the practice of empowering translation through artificial intelligence technology. Currently, large language models represented by GPT, through deep learning on massive amounts of data, have gained powerful semantic analysis capabilities, allowing them to generate meaningful text, which has strongly impacted the translation industry.
The research related to Natural Language Processing, which is the origin of machine translation, began around the 1950s. The early goal was to enable computers to understand, generate, and analyze human language, conducting tasks like syntax analysis, information retrieval, and sentiment analysis.In the field of translation, research in natural language processing mainly focuses on breaking down text into meaningful units and then tagging each word with the correct part of speech, understanding the deeper meaning of sentences by recognizing word meanings, semantic roles, and contextual implications.
After the development of natural language processing research, machine translation began to develop as an application branch, roughly divided into the following stages. The first to appear was rule-based machine translation, which relied on detailed dictionaries and grammar rules but was limited by the strictness of rules and the complexity of languages; it gradually evolved into statistical machine translation, which learned translations from large bilingual corpora by maximizing the probability of translation models to select the best translation; and then to neural machine translation, which uses deep neural networks to model the dependencies between all words in paired sequences, learning the dependencies between words in a sequence in an end-to-end manner, significantly improving translation quality.
Today, machine translation is gradually entering its fourth development stage—machine translation based on large language models. Large language models are a type of deep learning algorithm, typically trained on models with billions or more parameters, ultimately gaining the ability to understand and predict text. In recent years, generative artificial intelligence based on large language models has rapidly developed, enhancing the quality of machine translation while expanding the application boundaries of natural language processing. Large language models predict sequences of words given prompts, emerging the ability to predict more complex content, thus achieving multi-paragraph responsive translation and further improving translation quality.
The advantage of machine translation lies in its data learning capability, which can significantly improve translation efficiency.Traditional human translation requires translators to spend a lot of time researching relevant knowledge to ensure the accuracy of the text content.Human translators who want to attempt translations in different languages and fields need to invest additional learning costs and time, and they are also limited by their personal learning capabilities.Today’s large language models, through training on vast parallel corpora, have mastered the mapping rules from one language to another among massive knowledge, achieving large-scale automated real-time translation of texts, significantly shortening the translation cycle.At the same time, the application of pre-trained models and self-learning technology allows machine translation systems to adapt faster to new languages, new fields, and new scenarios, improving the flexibility and adaptability of translations.Moreover, with continuous technological advancements, the quality of machine translation has significantly improved.Currently, the accuracy of common sentence structures and terms in machine translation has reached over 80%, and the translation accuracy of some standardized texts (such as patent texts) even exceeds 95%.Additionally, through continuous learning and user feedback mechanisms, machine translation systems can continuously optimize and gradually overcome specific domain translation challenges.
Despite the significant progress made in machine translation based on large language models, there are still specific issues faced during the translation process. One is the lack of understanding of specific cultural backgrounds and professional knowledge. Machine translation models often lack a deep understanding of specific cultures, making it difficult to fully grasp the cultural implications in certain contexts. Cultural background knowledge is not only a vocabulary understanding issue but also involves social customs, historical backgrounds, emotional attitudes, etc. At the same time, in some highly specialized fields, machine translation lacks the accumulation of specific background knowledge and accurate terminology. Another issue is polysemy and commonsense judgment. Polysemous words have completely different meanings in different contexts, requiring reliance on commonsense for accurate judgment, and machine translation still falls short in this regard. Thirdly, there is the issue of logical coherence and consistency. Translation requires not only the correctness of words and sentences but also logical coherence and consistency. Sometimes, the text generated by machine translation lacks coherence and logical deduction between paragraphs, especially when dealing with complex causal or progressive relationships, machine translation may fail to maintain a clear logical structure, leading to disjointed content. Fourth, machine translation may be influenced by the training data when processing content, forming potential language or cultural biases. Some models may present biased expressions when dealing with topics such as gender, race, and religion, leading to discriminatory content in the translation output. This bias often reflects the cultural biases implicit in the training data, which can result in unfair translations or offend certain groups.
Although machine translation based on large language models has made significant progress, many issues remain. In the face of these issues, future development of machine translation needs to pay more attention to the depth and breadth of technological research and development, as well as collaborative cooperation with human translators.
Currently, machine translation is gradually moving towards a higher level of deep learning and artificial intelligence. Future technological development will focus more on optimizing models and improving algorithms to enhance translation accuracy, fluency, and naturalness. For instance, by introducing more complex neural network structures, attention mechanisms, and other technologies, further improving machine translation’s ability to capture hidden logical relationships and contextual information in language. Building and integrating specialized knowledge bases to enhance machine translation models’ capabilities in translating specialized terms and industry-specific expressions. Algorithm design should consider fairness and neutrality, avoiding the introduction of discriminatory biases during model training. Data processing should ensure diversity and representativeness to cover different languages and cultural backgrounds, reducing inaccuracies in translation caused by dataset biases. Meanwhile, new translation models such as cross-modal translation and personalized translation will also become important directions for technological research and development to meet the needs of different user groups.
Machine translation technology is not intended to replace human translators but to extend their capabilities. The development of machine translation based on large language models will promote a new mode of human-machine collaboration: machine translation and human translators forming a complementarity. As machine translation technology continues to advance, the role of human translators will also change; they will be liberated from tedious translation work and more involved in ensuring translation quality, conveying cultural backgrounds, and organizing writing logic in higher-level translation tasks. Additionally, translators can use machine translation technology as an auxiliary tool to improve translation efficiency and quality. For example, conducting preliminary translations through machine translation, followed by human refinement and proofreading, can enhance translation speed and accuracy.
With the development of technology, the translation industry is facing significant changes. Translators need to keenly capture these changes, timely adjust their translation strategies and methods, and improve their information literacy. Before translating, translators can use large language models to establish terminology databases and memory banks, accumulating specific domain corpora to help models improve professionalism and avoid mistranslations or biases in machine translation. During translation, translators can adopt a human-machine collaboration model, combining the translations generated by language models with human translations, leveraging the advantages of models in speed and diversity, while ensuring accuracy through human review. After translation, translators can utilize grammar and semantic analysis tools or other AI-assisted tools to automatically detect and provide feedback on the quality of translation results. Translators can also perform post-editing and human proofreading to ensure that translations meet the needs of usage scenarios, enhancing the overall consistency and accuracy of the text.
Looking to the future, as technology continues to advance and application areas become increasingly widespread, machine translation will play a more critical role in many fields, providing more convenient and efficient tools and methods for cross-language communication.
This article is a phased achievement of the major project of the National Social Science Fund “Research on the Organization and Innovative Writing of the Century-Long History of Chinese News Communication” (22&ZD321)
The author is a professor at the School of English, Xi’an International Studies University
For communication, please contact us
Email: [email protected]
