Machine translation (MT) technologies have significantly advanced, offering solutions that are not only efficient but also remarkably accurate. DeepL, in particular, has garnered attention for its ability to deliver translations with unparalleled precision and fluency. This article aims to shed light on DeepL’s performance by comparing it with other major MT engines, drawing insights from a recent comprehensive study. Furthermore, I’ll dive into the practical applications of the DeepL Translator API.

Machine Translation Analysis and Comparison

The advancement of machine translation (MT) technologies has been a game-changer for global communication, offering businesses, content creators, and linguists the ability to translate vast amounts of text quickly and with increasing accuracy. Among the various players in the MT field, DeepL has emerged as a particularly noteworthy contender, renowned for its superior translation quality. In this segment, we’ll delve into a comparative analysis of DeepL against other leading MT engines, drawing on insights from a comprehensive study detailed in a recent report.

Source: “The State of Machine Translation for Websites,” a study brought to you by Nimdzi Insights and Weglot, 2022.

Comparative Framework

The study evaluated five leading MT technologies across seven language pairs, focusing on aspects such as overall usability, accuracy, reliability, and the editing effort required post-translation. The MT engines included in the study were Amazon Translate, DeepL, Google Cloud, Microsoft Translator, and ModernMT. The evaluation was carried out by professional linguists who reviewed translations for content extracted from a service company’s corporate website, ensuring a practical and relevant assessment framework.

Key Findings

• Overall Usability: DeepL, alongside other engines, showed high overall usability for translating company website content. However, DeepL often led the pack, with translations that required minimal post-editing to be deemed usable in a professional context.
• Editing Effort: One of the standout insights from the report was the reduced editing effort required for DeepL-translated content. DeepL translations generally required fewer corrections compared to its competitors, indicating a closer initial alignment with the expected quality standards of professional linguists.
• Accuracy and Reliability: Across the board, MT technologies have made significant strides in terms of accuracy and reliability. Yet, DeepL frequently received higher ratings in these categories, suggesting that its algorithms may be better tuned to handle the nuances of language and context more effectively than others.
• Handling of Complexities: The study highlighted specific challenges encountered by MT engines, such as translating content out of context and managing language-specific nuances (e.g., differentiating between European and Brazilian Portuguese). DeepL’s performance in these areas was noteworthy, often outperforming other engines by producing translations that were both contextually accurate and stylistically appropriate.

Comparative Analysis Highlights

• No-Touch Translations: DeepL demonstrated a higher percentage of “no-touch” translations, meaning that a significant portion of its output was used as-is, without any need for edits. This indicates not only efficiency but also a high level of trust in the quality of DeepL’s translations.
• Engine Performance: While each MT engine had its strengths, DeepL consistently ranked at or near the top in several key performance metrics. For example, DeepL had one of the lowest edit distances across all evaluated language pairs, suggesting that its translations were closer to the final desired output from the outset.

The comparative analysis of MT technologies reveals DeepL as a frontrunner in the field, offering high-quality translations that stand out in terms of usability, accuracy, and reliability. While the landscape of MT is continuously evolving, with each engine bringing its own set of advantages, DeepL’s performance in this study underscores its potential as a powerful tool for anyone needing professional-grade translations at scale.

In the following section, we’ll explore how to harness DeepL’s capabilities through its API, unlocking the full potential of this advanced translation technology for your projects and workflows.

The Translator engine: from GPT-4 to DeepL API

Recently, I embarked on creating a translation platform utilizing GPT-4, attracted by its advanced language processing capabilities. While the platform demonstrated promising quality in some languages, it revealed stark inconsistencies in others—ranging from commendable accuracy to notably poor translations.

This experience underscored the need for a model specifically tailored to translation tasks, leading me to DeepL, and its renowned precision across various languages made it an ideal choice for ensuring consistent quality in translations.

How to Get a DeepL API Key

1. Create a DeepL account: Go to the DeepL website (https://www.deepl.com/pro-api) and sign up for an account.
2. Choose a Plan: You’ll need either the DeepL API Free or the DeepL API Pro plan.
3. Find Your Key: Once your plan is set up, go to your DeepL account settings and find the section labeled “API Keys”. Here you’ll be able to find your existing key or generate new ones.

Costs

• DeepL API Free:
  • Translates up to 500,000 characters per month for free.
  • Limited features compared to the Pro plan.
• DeepL API Pro:
  • Unlimited translation volume.
  • Usage-based pricing. You are charged per million characters translated (currently $25.00 USD per 1,000,000 characters).
  • Prioritized requests, greater data security, and the ability to create custom glossaries.

Important Notes

• The API keys and API domains are different for the Free and Pro plans. Make sure you use the correct ones.
• The DeepL API does not grant access to the DeepL web translator or the desktop applications.

DeepL API Guide

In this section, we’re going to dissect and understand all the request body parameters available in the DeepL API. These parameters are crucial for tailoring translation requests to fit specific needs and scenarios, ensuring that the output aligns perfectly with your requirements.

• text (required)
  • Description: The text segments you want to translate. Supports multiple segments in one request.
  • Type: array[string]
• source_lang (optional)
  • Description: The language of the text being translated. If not specified, the API attempts to auto-detect the language.
  • Type: string
• target_lang (required)
  • Description: The language into which you want the text to be translated.
  • Type: string
• context (optional)
  • Description: Additional context to improve translation quality, not translated itself.
  • Type: string
• split_sentences (optional)
  • Description: Controls how the engine splits the text into sentences, with options for no splitting, splitting by punctuation and newlines, or by punctuation only.
  • Type: string
• preserve_formatting (optional)
  • Description: Instructs the engine to maintain the original formatting of the text.
  • Type: boolean
• formality (optional)
  • Description: Sets the desired level of formality in the translated text, applicable to specific target languages.
  • Type: string
• glossary_id (optional)
  • Description: Specifies a glossary to apply to the translation, requiring matching source and target languages.
  • Type: string
• tag_handling (optional)
  • Description: Specifies how to handle XML or HTML tags in the text, ensuring proper translation and retention of markup.
  • Type: string
• outline_detection (optional)
  • Description: Enables or disables automatic detection of XML structure, allowing for manual control over structural elements.
  • Type: boolean
• non_splitting_tags (optional)
  • Description: Tags that should not cause sentence splits, useful for preserving specific structures in XML/HTML content.
  • Type: array[string]
• splitting_tags (optional)
  • Description: Tags that always cause sentence splits, aiding in the correct segmentation of structured documents.
  • Type: array[string]
• ignore_tags (optional)
  • Description: Tags indicating sections of text that should not be translated, preserving parts of the content as-is.
  • Type: array[string]

Enhancing Quality: Migrating Translator App from GPT-4 to DeepL (python)

Moving from the GPT-4 model to the DeepL API was surprisingly seamless. With GPT-4, crafting detailed prompts outlining the system’s role and desired output format was necessary. DeepL eliminates this complexity; its API is focused purely on translation. Providing the text and target language is all that’s required. This direct approach simplifies integration and saves development time often spent constructing those elaborate GPT-4 prompts.

The deepl library is a Python package that provides a convenient interface to interact with the DeepL translation API. It simplifies the process of sending text to DeepL’s servers and receiving accurate translations in return.

• Brings in Code: The import deepl statement essentially loads all the code contained within the deepl library into your current Python script. This gives you access to the functions, classes, and other tools provided by the library.
• Creates a Namespace: When you import a library, a namespace is created. In this case, all the components of the deepl library are placed within the deepl namespace. To use these components, you’ll usually prefix them with deepl., (e.g., deepl.Translator).

A Basic DeepL API Code Example

• Installation: You’ll need to install the deepl library using pip before you can import it: pip install deepl
• API Key: You’ll need a DeepL API key to use the library (https://www.deepl.com/pro-api).

Translator: This class is the core of the library. You create a Translator object using your DeepL API key, and then use it to perform translations:

translator = deepl.Translator("YOUR_API_KEY")
result = translator.translate_text("Hello world", target_lang="DE")
print(result.text)  # Output: Hallo Welt

In Conclusion:

The move from the more general language capabilities of GPT-4 to the DeepL API demonstrates a clear prioritization of translation accuracy and specialization. While GPT-4 offers impressive language generation and understanding, DeepL’s focus on translation delivers superior results for this specific task. This shift highlights the importance of selecting the right tool for the job.