essay-Machine Translation
Machine translation has always been a raised question. Dating back to 1954, IBM released 701, which translated based on rules, showing the prospect of that direction. In the 1990s, the statistical model superseded the rule-based model. In 2016, Google Translation introduced NMT, which uses Neural Networks and has a close relation with LLM. Their techniques intertwine and improve each other. Nowadays, the optimism on machine translation is generally felt--which is apt in some ways. They perform in benchmarks like MQM, competent even compared with multilinguals on translation. Some people are struck by how human improvement seems less and less required except for the "moral crumple zone", or when responsibility needs someone to take. They predict that most humans in translation would be squeezed out.
Machine translation has long been a topic of interest and inqury. Dating back to 1954, IBM introduced the IBM 701, a rule-based translation system that demonstrated the potential of computational linguisitic. By the 1990s, statistical models began to supersede rule-based systems, marking a significant shift in the field. In 2016, Google Translation introduced Neural Machines Translation (NMT), which uses Neural Network and is closely linked to large language model(LLM). These techniques have evolved together, enhancing each other's performance.
Today, there is widespread optimism about machine translation, and it feels apt. MT systems perform well on benchmarks like MQM and can even rival multilingual human translators in specific tasks. Some obserevers are struck by the decreasing need for human involvement in translation, except in "moral crumple zone"-- a face to take the blame when things go wrong, but with no reasonable expectation of improving outcomes. As a result, they predict that human translators will be squeezed out of their roles.
That may be complacent. Powerful as it is, machine translation still leaves problems, many of which draw similarities with those in LLM, like weakness in memory, reasoning, and planning. People who rely heavily on machine translation may often find the machine can't remember the mentioned entity's name, thus often translating the same thing with different names in an article. They also sometimes hallucinate components, like in translating language where parts like subjects are often omitted like Japanese, therefore contorting the original meaning. And idioms or new trend expressions are often translated literally. They find it hard to balance between fidelity and transparency.
However, such prediction may be overly complacent. Despite its power, machine translation still have notable limitations, many of which are similar to challenges faced by LLMs. For instance, MT systems struggle with memory, reasoning, and planning. Users relying heavily on MT may find out that machines fail to maintain consistency, like translating the same entity's name differently within a single passage. Additionally, they sometimes hallucinate components, esecially in language like Japanese where elements such as subjects are frequently omitted. This can contort the original meaning. MT systems also tend to translate idioms and contemporary expressions literally, failing to capture their intended implication. Striking a balance between fidelity and transparency remains a persistent challenge.
Translators sometimes need implicit information outside the text. For example, translating ”I like you“ into Japanese, may need context like who said that to whom to avoid rude expressions like "あなた". DeepL, an outstanding company in translation, is dedicated to providing corporations with high-quality service. Its founder said that translation needs to take into account things like style, and the machine may need the be told about those things to generate an appropriate answer. DeepL provides different styles of translation and is still improving its performance in specific fields like contact, trying to avoid impolite, hard, or casual expressions.
Implicit information outside the text may be helpful for machine translation. For example, translating "I like you" into Japanese requires context to avoid rude expressions like あなた. DeepL, a prominent MT company, excels in providing high-quality-services to buinesses. Its founder has noted that MT requires considering elements like style, and machine may need to be explictly needed instructed to follow. Deepl thus offer style-specific translations and continue to imporve its performance in specialized doomains like contacts, striving to avoid impolite, overly casual, or odd expressions.
User-oriented customize may be a way out and a necessity. DeepL cooperates with companies to use their internal material to train and optimize translation models and learn about customized expressions, jargon as well as culture.
Customizability may be key to and necessary for improving MT. DeepL cooperates with companies to learn from, as well as train and optimize translation models using internal materials, enabling the system to optimize learn customized expressions , jargon and cultural nuances.
Is learning languages becoming meaningless? For people who just want to use it for travel or simple use, it stands so. But learning languages is a good humanizing process, in which you can also improve your brain. The sense of achievement of gradual understanding and proficiency can also be tempting. Moreover, language is representative of a certain culture and way of thinking. It is a way of people contacting each other, with implications and vague parts that are elusive (at least now) to translators. Using a foreign language to communicate is establishing and maintaining a specific connection, and the relationship is best without an intermediate.
Does the rise of MT render language learning meaningless? For those who only need basic communication for travel or simple tasks, this may be true. However, learning a language offers intrinsic value. It fosters cognitive growth and provides a profound sense of achievement as proficiency develops. Language is deeply tied to culture and thought patterns, serving as a medium for human connection rich with implications and subtleties that elude machines nowadays. Communicating direclty in a foreign language establishes and strengthens relationship without the need for intermediaries, making it a uniquely meaningful connection and experience.
a face to take the blame when things go wrong, but with no reasonable expectation of improving outcomes.
浙公网安备 33010602011771号