The Evolving Quality of Machine Translation

The days of clunky, awkward translations are behind us—thanks to the machine translation (MT) revolution.

Researches continuously show that a vast majority of internet users prefer to access information in their native language (65% worldwide, and 90% in the EU). This shows that the need for website translations is skyrocketing. Luckily, MT has come a long way since its humble origins, making it possible to fulfill this need.

Automated machine translation is no longer just for enterprises, governments, and global organizations. Today, as technology advances and becomes more affordable, SMEs and start-ups can benefit from the ever-improving translation quality of MT to break down language barriers and foster international collaboration and reach new markets.

In this article, we’ll dive into how machine translation systems have improved in quality over the years.

The Quality of Early Machine Translation

The early days of machine translation (MT) development were exciting, but it quickly became clear that the system was limited. Initial approaches relied on either rule-based or statistical models.

As the name suggests, rule-based systems used linguistic rules and dictionaries to translate words and phrases. However, languages are complex and ever-changing. So, manually creating rules for every grammatical structure and vocabulary nuance turned out to be an impossible task.

Statistical machine translation (SMT), on the other hand, analyzed tons of existing translated texts to identify patterns and probabilities. This approach was easier, but it created literal translations that lacked the natural flow and context awareness of human translations.

These early MT systems frequently produced awkward translations, struggling with idioms, metaphors, and cultural references. Only a small percentage of early machine-translated sentences were considered “understandable” and “accurate.”

Early MT also had trouble distinguishing between the various meanings of a word based on its context.

For example, the word “bank” could be a financial institution or the side of a river, but early MT systems failed to differentiate. Grammatical errors were also common, as early MT systems struggled with complex sentence structures and word order variations.

The Neural Network Revolution

After many decades of unreliable translations, neural machine translation (NMT) finally changed the game in the mid-2010s.

Unlike its predecessors, NMT uses artificial neural networks to provide higher-quality translations. Google’s 2016 research* shows that NMT outperformed traditional SMT systems by a huge margin, reducing translation errors by 60%.

NMT heavily relies on word embeddings, which represent words as vectors in a high-dimensional space. They create semantic relationships between words, allowing NMT models to understand the meaning of words in context rather than simply just their surface forms.

The encoder-decoder architecture is another key component of NMT. The encoder reads the source sentence and converts it into a mathematical representation. Then, the decoder takes the encoded representation and translates it into the target language, word by word.

Attention mechanisms allow the decoder to focus on relevant parts of the input sentence at each step of the translation process. As a result, NMT systems create more accurate and contextually correct translations.

NMT has achieved what earlier MT systems couldn’t. Since it captures semantic meaning through word embeddings, NMT can also handle ambiguous words and phrases. That means it can also translate complex sentences fluently and coherently.

Today, popular translation tools and services like Google Translate, DeepL and Microsoft Translator utilize NMT. These tools can serve all kinds of linguistic purposes, from personal translation use to business cases like website translation.

Context: The Key to Understanding

Language isn’t just a collection of words – the meaning comes from the context. Without context, we can’t piece together the full picture. The same goes for machine translation.

The transformer models, like Google’s BERT and OpenAI’s GPT, use self-attention mechanisms, which allow them to weigh the importance of different words in a sentence based on their relationship to other words.

Meanwhile, early MT models had limited context windows, meaning they could only consider a few words at a time. Transformer models have much larger context windows. They can analyze entire sentences or even paragraphs to understand the meaning.

This is mainly due to the increase in context window tokens. For instance, Google’s Gemini 1.5 went from 128,000 tokens to 1 million tokens* – the longest context window for any MT model yet.

How Context Improves MT

Context impacts multiple crucial aspects of machine translation:

• Disambiguation: Context helps MT systems determine the correct meaning of polysemous words (words with multiple meanings) based on how they are used in a sentence.
• Pronoun Resolution: Context allows MT systems to identify the referents of pronouns like “he,” “she,” or “it.”
• Idiomatic Expressions: With context, MT systems can translate idiomatic expressions and figurative language that might not make sense in literal translations.

Data-Driven Model Training for AI Translation

MT models feed on massive amounts of data to better understand and translate language. This is why large-scale, diverse datasets are crucial for modern MT development.

MT models trained on larger datasets can understand complex sentences, idiomas, and domain-specific terminology like never before. Plus, exposure to diverse language patterns allows MT models to generate more natural-sounding translations that flow smoothly.

Large volumes of high-quality training data are obtained in three ways:

• Parallel Corpora: Collections of texts that have been translated by humans into multiple languages.
• Web Crawling: Collecting text data from the internet, including websites, blogs, and social media, with the help of a bot.
• Back-Translation: Translating a text into another language and then back into the original language. This creates synthetic parallel data that can be used to augment existing datasets.

Organizations like Common Crawl* and the Linguistic Data Consortium* play a huge role in collecting large-scale multilingual datasets for MT research and development.

How MT Quality Has Improved Over the Years

Let’s compare early machine translations with neural machine translation (NMT) models to see how far MT has come.

Example 1: English to French

“The spirit is willing, but the flesh is weak.”

Early MT: “L’esprit est disposé, mais la chair est faible."

Modern NMT: “L’esprit est fort, mais la chair est faible.”

Modern NMT correctly identifies the English idiom and translates it naturally. Meanwhile, early MT is grammatically correct but too literal, so it misses the cultural nuance.

Example 2: English to Chinese

“He kicked the bucket.”

Early MT: “他踢了桶.”

Modern NMT: “他去世了.”

Modern NMT understands that “kicked the bucket” is an idiom for dying and translates it correctly. Meanwhile, the early MT model’s literal translation is meaningless in Chinese.

Example 3: English to German

“The man who was walking down the street saw a cat that was chasing a mouse.”

Early MT: “Der Mann, der die Straße entlangging, sah eine Katze, die eine Maus jagte.”

Modern NMT: “Der Mann, der die Straße entlangging, sah eine Katze, die hinter einer Maus her war.”

Modern NMT produces a more concise and natural-sounding translation, while the early MT output is overly complex.

Research on the Evolving Quality of Machine Translation

A 2021 study by the Moscow Region State University* quantifies exactly how far machine translation has come in the last 50 years. Google’s machine translation system has seen a noticeable improvement over the past five years, with a 4% increase in accuracy. Similarly, the PROMT system, which only translates scientific texts, has shown a remarkable 14% improvement in translation quality since 2016.

Both systems have shown over 75% accuracy for full-sentence matches. This level of accuracy is unprecedented, as sentences with higher match percentages (70-100%) contain only minor errors.

However, NMT still requires some work. A research at University of Cambridge* shows that larger NMT models now require smaller batches of data, resulting in slower training.

Training these models for large vocabularies is also tricky due to Zipf’s law, which states that most words in a language are rare. That means it’s hard to train robust word embeddings (the foundation of NMT) for infrequent words.

The good news is that modern word-based NMT models address this by limiting vocabulary to the most frequent words and replacing others with a UNK token.

Plus, NMT is still a work in progress, and researchers are hard at work making these models more accurate and less error-prone. Considering how far this type of machine translation has come in just the past decade, the future looks bright.

Conclusion

Early machine translations were groundbreaking for their time, but they’re nothing compared to the near-perfect models we have today. Sure, they have their limitations, too, but machine translation has come a long way since its origins.

As this technology continues to advance, businesses have the chance to connect and communicate on a global scale. With website translation services like Easyling, catering to international customers has become easier than ever.

Book a demo with Easyling today and discover how easy it is to make your content accessible to the world.