MTPE is OUT, Augmented Translation is IN

Machine Translation Post Editing: Impacts on Translation Agencies

Understanding Machine Translation (MT)

• Gabriel Fairman introduces the topic of machine translation post-editing (MTPE) and its financial implications for translation agencies.

• Machine translation is defined as processing text through various engines (statistical, rule-based, neural), producing outputs without human interference.

• Tuning an engine involves adjusting rules and algorithms, while training it refers to the data fed into the system.

Defining Post Editing (PE)

• Post editing refers to the human process of reviewing and adjusting machine-translated texts to ensure they are viable translations.

• The definition of post editing varies; some see it as perfecting translations, while others view it as correcting blatant errors only.

• For this discussion, MTPE encompasses any edits made after machine translation.

Efficiency vs. Challenges in MTPE

• MTPE promises enhanced efficiency; reviewing a pre-translated document can be quicker than translating from scratch.

• However, MTPE can be complex and time-consuming due to hidden errors that may not be immediately apparent in machine-generated text.

• The effectiveness of MTPE varies by subject matter; performance may differ significantly between fields like finance and marketing.

Variability in Performance

• High-resource languages may yield better results in MT than low-resource languages due to differences in training data availability.

• The promise of greater efficiency is often undermined by inconsistent quality across different translations.

Human Interpretation and Standardization Issues

• MTP processes are subjective; instructions for editing can be interpreted differently by individuals, complicating standardization efforts.

• This variability leads to challenges in ensuring consistent quality across translated documents despite clear guidelines.

Financial Implications for Translation Agencies

• Many agencies began using MT for tasks previously handled by human translators, often paying less but requiring more cognitive effort from editors.

Understanding Augmented Translation

The Flaws of Traditional Machine Translation

• The speaker critiques traditional machine translation (MTP), describing it as a poorly adopted technology that often prioritizes cost-cutting over quality.

• Augmented translations are introduced as an alternative, functioning similarly to MTP but with significant differences in implementation and effectiveness.

Key Features of Augmented Translations

• Context-sensitive translation is highlighted, which utilizes various context sources (e.g., Microsoft’s machine translation, glossaries) to enhance accuracy.

• Unlike MTP, augmented translations maintain a continuous dialogue between the translator and the engine, allowing for real-time adjustments and learning from user input.

Advantages Over Machine Translation Post Editing

• The speaker expresses frustration with repetitive changes in MTP post-editing, emphasizing the inefficiency of having to repeatedly correct the same errors.

• In contrast, context-sensitive translation learns from each change made by the translator, improving future suggestions based on previous interactions.

Enhanced Semantic Verification

• Augmented translations provide additional semantic verifications that can identify issues like mistranslations or cultural misappropriations—areas where traditional QA checks may fail.

• Edit distance is introduced as a measure of effort in editing; lower edit distances indicate less work needed to align machine-generated text with human preferences.

Overall Experience and Future Potential

• The speaker notes that augmented translations allow for a more enjoyable experience compared to MTP due to their adaptability and contextual understanding.

Understanding the Relationship Between Edit Distance, Cognitive Effort, and Translation Quality

The Complexity of Edit Distance and Cognitive Effort

• Connecting edit distance to cognitive effort is challenging; it's even harder to link cognitive effort to translation quality.

• A minimalistic editing approach may result in low edit distance and cognitive effort but can yield high translation quality.

• Conversely, a high edit distance occurs when one rewrites everything from scratch, leading to increased cognitive effort and potential introduction of errors.

Need for Research in Translation Metrics

• Establishing a fair framework for evaluating translation requires extensive research; current industry standards are lacking.

• There is flexibility in setting parameters for net rates in machine translation (MT), but common denominators for payment structures remain undefined.

Challenges with Machine Translation Confirmation

• The criteria for confirming machine translations need clarity; questions arise about what percentage of text warrants additional payment.

• Each confirmation not only validates segments but also contributes to training the linguistic corpus as a whole.

Distinction Between MTPE and Augmented Translations

• There is a significant difference between machine translation post-editing (MTPE) and augmented translations; the latter offers a more sophisticated, organic experience.