Expert Introduction
Feng Zhiwei, researcher at the Language Application Research Institute of the Ministry of Education, doctoral supervisor, member of the academic committee, computational linguist, specializes in interdisciplinary research in linguistics and computer science. Currently, he is a part-time professor at Peking University, Zhejiang University, and Dalian Maritime University, and a Tianshan Scholar at Xinjiang University. He developed the world’s first machine translation system from Chinese to multiple foreign languages, the world’s first Chinese terminology database, and has received the Austria Vist Award and the NLPCC Outstanding Contribution Award from the Chinese Computer Society. He has published over 500 papers in Chinese and foreign languages and authored more than 40 works including “Introduction to Modern Terminology”, “The Theory and Methods of Formal Analysis in Natural Language Computing”, “Overview of Natural Language Processing”, and “Handbook of Formal Analysis in Natural Language Processing”.
(Image Source: Visual China)
Q: What has been the development process of computational linguistics in our country?
Feng Zhiwei: I believe the research history of computational linguistics in our country can be divided into three main periods.
First is the embryonic period. In 1954, the first Russian-English machine translation system was developed in the United States. Chinese researchers began to realize that mathematical models in linguistics have research value. In 1956, the national scientific planning proposed to carry out machine translation and formal research on natural language. In 1959, our country developed the first machine translation model, capable of translating Russian into Chinese, achieving a breakthrough from nothing to something.
Next is the recovery period. From 1976 to the late 1980s, China actively engaged in international exchanges and learning, and computational linguistics gradually evolved from a technical issue into a discipline. Around 1982, Chinese scholars began to appear at top academic conferences in computational linguistics.
Then comes the development period. After 1989, machine translation began to move towards industrial applications, with companies providing translation products. This period emphasized resource construction, such as data collection and establishing bilingual corpora. After 2016, the accuracy of neural machine translation can reach over 95%, and machine translation gradually became practical, with domestic companies launching machine translation systems.
Q: What are your views on neural machine translation?
Feng Zhiwei: The deep learning technology used in neural machine translation is essentially a pattern recognition technology. When computers translate, they simply use an encoder-decoder model to convert text data and do not truly understand the content of the text. However, this is not entirely equivalent to “parroting”, because relying on large-scale corpora and powerful computational capabilities, deep learning can convert the semantics within the language into contextual relevance and describe such relevance using word vectors. Computers can understand some local content within language texts through deep learning and delve into the complex relationships among various parameters of language data, thus possessing a strong ability to describe internal language relationships.
Although neural machine translation performs well, it often struggles with polysemy, determining referential relationships, and lacks interpretability, as deep learning and neural network technologies are based on pattern recognition and are connectionist approaches based on big data. We do not fully understand the operational mechanisms of neural machine translation, and its practical applications in translation should be approached with caution.
Q: What challenges do you think neural machine translation currently faces?
Feng Zhiwei: As an important branch of natural language processing, neural machine translation has achieved an accuracy of over 98% in some fields and languages. However, I believe neural machine translation does not possess true human intelligence and faces challenges such as the inability to understand emotions and a lack of common sense.
Human understanding of natural language relies not only on various internal relationships of the language but also on background knowledge from the external physical world, spiritual world, and social historical world. Each symbol in natural language text, each structured string of symbols that adheres to rules, has complex connections with the external objective world in the human brain. These complex connections are represented not only in conceptual symbolic forms but also have visual, auditory, tactile, and bodily representations, and even deeper psychological and emotional representations as well as social and cultural backgrounds. Currently, deep learning has developed certain visual and auditory capabilities through multimodal modeling, but it still lacks the ability to process tactile and bodily sensations, nor can it handle the rich and diverse common sense outside of language. Therefore, deep learning cannot yet explore the various complex connections between language data and the external world, which is the fundamental difference between current artificial intelligence and human intelligence.
Translation is a high-level intelligent activity of humans, involving not only the internal structure of language but also a myriad of complex and rich elements such as everyday knowledge, social knowledge, historical knowledge, and cultural background knowledge from outside the language. These non-linguistic elements constitute the “humanity core” of translation. Currently, although neural machine translation has made significant progress and has some ability to simulate the internal structure of human language, its capacity to simulate the external world and social historical background is still very limited, and it struggles to handle these complex and rich “humanity cores”. Therefore, when neural machine translation encounters “humanity core”, it often finds itself in a dilemma.
Q: How do you view the relationship between machine translation and human translators?
Feng Zhiwei: In recent years, translation technology has developed rapidly, and some believe that translators will be out of jobs. However, I think the translation capabilities of machine translation have been overstated; it cannot replace human translators, as complex high-end translation work must be undertaken by humans. On one hand, translations of literary works, scientific literature, and other specialized fields still need to be completed by humans; on the other hand, translation tasks with high confidentiality requirements, as well as simultaneous and consecutive interpretations in important situations, also require human involvement. Specifically, literary translation requires translators to possess a high level of humanistic literacy and a deep understanding of the cultural background of the source language, as well as the ability to skillfully and creatively use the target language, which is difficult for machine translation to achieve and must be done by human translators. In scientific translation, polysemous terms can represent various concepts in different fields, and machine translation often struggles to correctly identify such terms, leading to translation errors that require human judgment. Furthermore, although machine translation can perform simultaneous and consecutive interpretation, in real-time translation scenarios, machine translation often struggles to correct errors in a timely manner, which may lead to irreparable consequences, thus important simultaneous and consecutive interpretations must also be handled by human translators. Therefore, it is clear that machine translation cannot replace human translators; high-end translation experts are irreplaceable by machine translation.
I believe machine translation will become a good friend and a capable assistant to human translators. The two should coexist harmoniously and complement each other. In the era of artificial intelligence, the intelligence levels of various translation technology tools are increasing, which will help enhance the translation efficiency of human translators. Translators should keep pace with the times, embrace technology, learn technology, and master technology.
Q: What impact do you think large language models have on the translation industry?
Feng Zhiwei: I believe large language models (LLMs) provide new opportunities for the translation industry while also posing new challenges. LLMs use machine learning and natural language processing technologies to achieve automatic translation, making translation faster, more convenient, and accurate, thus saving time and economic costs; LLMs can provide customized translations based on user needs and preferences, improving translation quality and user experience, offering personalized services; LLMs can help enterprises engage in more intelligent communication with customers, enhancing customer satisfaction and loyalty; LLMs make communication between different languages and cultures easier, promoting globalization and cross-cultural exchange; LLMs can collect large amounts of language data and analyze and mine this data to generate valuable business insights. In summary, LLMs will drive intelligent reform and efficient innovative development in the translation industry through technological innovation.
Of course, LLMs also bring new challenges to the translation industry. With their translation capabilities, LLMs can generate a large number of translation results in a short time, making them a low-cost and high-efficiency alternative to human translators; at the same time, their translation results can automatically incorporate data information from the large model, with significantly improved contextual understanding compared to neural machine translation. Therefore, LLMs, which combine contextual understanding and text refinement functions, have a huge impact and shock on the translation industry, leading to a gradual decline in market share for some traditional translation companies. As LLM technology continues to develop, more and more enterprises will begin to use LLMs to enhance the quality and efficiency of their translation products, thus those companies that cannot provide higher quality services will face the risk of exiting the market.
To address these challenges, the translation industry can maintain market competitiveness by strengthening its core competencies and expanding into new fields. At the same time, it can also consider integrating with LLM technology to improve service quality and efficiency.
Q: How do you foresee the development and application of large language models in 2024? What impact will it have on translation and international communication?
Feng Zhiwei: Large language models have profoundly changed the way translation knowledge is produced in the past, shifting the research subjects of translation from individual studies to collaborative group intelligence, the research process from experience accumulation to data analysis, and the research form from a single discipline to multidisciplinary, from text or speech data to multimodal data. This represents a significant methodological change and major innovation in the paradigm of translation knowledge production, which will promote the development of translation and international communication work.
Large language models are a transformative artificial intelligence technology that will reshape social and technological development, but they also carry various obvious and foreseeable risks.
First, due to their inherent “hallucination” problem, large language models may generate content that is unrealistic, inconsistent, or does not meet human expectations, potentially containing discriminatory, biased, and sensitive information. They may also propagate harmful information and toxic content from their training data, generating misleading and false information.
Second, large language models may be exploited by malicious actors to perform harmful actions. They could generate realistic fake news or assist hackers in attacking devices on the internet. These malicious actions can negatively impact our daily lives and even cause serious harm to society as a whole. As their capabilities increase, large language models may also exhibit goals such as self-preservation, self-enhancement, and resource acquisition, which could be pursued by nearly all digital agents.
Additionally, training and deploying large language models require substantial computational resources and electricity, incurring huge investments and social wealth.
These are the disadvantages and risks that large language models may produce. Multilingual digital intelligent agents based on large language models will increasingly integrate into our daily lives. To overcome these drawbacks and avoid various unpredictable risks, it is necessary to promote research on alignment technologies for large language models, ensuring that their outputs and behaviors align with human expectations and values. Any misaligned behavior could lead to unforeseen serious consequences.
In 2024, alignment of large language models should become an important aspect of language governance. In language governance, we should establish a testing platform for alignment methods of large language models to serve as a platform for testing alignment method experiments and proposals, which will help formulate more stable alignment methods and reach consensus on key issues, and establish a consistent scientific framework for the alignment of large language models. These are the issues that language governance should focus on in the era of large language models.
Since 2023, a series of governance measures and initiatives for large language models have been proposed both domestically and internationally.
On August 15, 2023, our country officially implemented the “Management Measures for Generative Artificial Intelligence Services”, which stipulates the institutional requirements for service providers and points the way for the future development of the generative artificial intelligence industry.
On October 18, the Central Cyberspace Affairs Commission released the “Global Artificial Intelligence Governance Initiative”. The initiative proposes that the development of artificial intelligence should adhere to the principles of mutual respect and equality, allowing all countries, regardless of size or strength, to have equal rights to develop and utilize artificial intelligence. Meanwhile, the National New Generation Artificial Intelligence Governance Professional Committee of China formulated the “Ethical Norms for the New Generation of Artificial Intelligence”, and the Ministry of Foreign Affairs of China drafted the “Position Paper on Strengthening the Ethical Governance of Artificial Intelligence”.
On November 1, the first Global Artificial Intelligence Safety Summit was held in the UK, where the “Bletchley Declaration” was released, pointing out that many risks associated with artificial intelligence are fundamentally international and should be addressed through international cooperation.
On December 8, the European Parliament, EU member states, and the European Commission reached an agreement on the “Artificial Intelligence Act”, which will become the world’s first comprehensive regulatory legislation in the field of artificial intelligence.
On December 28, OpenEval platform, China Software Testing Center, and other organizations jointly released the “2023 White Paper on Benchmark Testing of Artificial Intelligence Large Models”, indicating potential security risks of large language models and emphasizing the importance of value alignment as their capabilities evolve.
Governance of large language models has become a global consensus, which is a positive development.
Translation and international communication work should pay attention to the governance of large language models to ensure the healthy development of translation and international communication efforts.
Attention! Good News! Reply “Search in the Account” in the WeChat official account dialog box to get the link for quick search of any information in the “Translate World” account~
Recommended Reading
How to say “person-times” in English during the Spring Festival travel rush?
“Starting Work with Good Luck” is not the same as “happy start working”! A more authentic expression!
Recruitment | China Daily New Media Center’s Food Column is looking for interns
How to say “Year of the Dragon” in English? Dragon or Loong?
Translation | How to say “hard seat”, “hard sleeper”, “second-class seat”, “business class seat” on the train in English?
