Beyond the hype: Understanding the true value of Machine Translation Post-Editing (MTPE)
2024-02-27 | Malina Prusinska-Kurpiewska
Differentiating Proofreading, Editing, and MTPE Scopes
To unlock its full potential and ensure optimal content, whether text is created by humans or machines, proofreading/editing is necessary. While proofreading/editing are terms commonly associated with human-generated translations, the refinement of machine-translated text is termed Machine Translation Post-Editing (MTPE). Let’s explore the definitions and scopes of proofreading and editing to ensure an accurate representation and description of MTPE services later on.
Proofreading
Proofreading is the process of carefully reviewing a text to detect and correct mistakes related to grammar, punctuation, and spelling. The scope of the work is limited to these aspects, and no edits are made to correct style or general readability.
As a result, proofreading does not necessarily require reading the final text against the source; in extreme, yet undesirable cases, proofreaders may not even be familiar with the source language.
The primary goal is to ensure the target text is grammatically correct and free of mistakes such as typos. Proofreading occurs after the translation is complete and is typically the final step before returning the text to those who have ordered the translation.
Editing
Editing involves a more comprehensive review of the text, encompassing all the checks typically conducted by a proofreader (grammar, spelling, punctuation). Additionally, it addresses issues related to style, coherence, clarity, accuracy, precision of translation, cultural adaptation, and overall structure.
Editors not only correct linguistic errors but also rephrase sentences to enhance flow, improve the organization of ideas, and ensure that the target text conveys the intended message with the intended tone, aligning with the target audience.
Due to these considerations, editors have to read the target text against the source. While editing may be the final step in the localization process, it can also commence while the translation is still in progress. This collaborative approach between the translator and editor helps refine the initial draft, saving time and elevating the final quality of the text.
At LocalizeDirect, we rarely perform just proofreading; the so-called second pair of eyes delivers the final product, which is both proofread and edited.
MTPE
Machine Translation Post-Editing (MTPE) involves refining machine-generated translations to eliminate obvious mistakes related to both grammar and translation accuracy. In scope, it may be positioned somewhere between traditional proofreading and editing: MTPE involves more than just correcting grammar, syntax, punctuation, or spelling but is not as in-depth as full-on editing.
Understanding MTPE Borders
While the scope of MTPE is still being discussed and may vary from vendor to vendor, the most popular approach follows the borderlines listed below.
What’s Included in MTPE Work
The focus of post-editors in MTPE spans various aspects, including grammatical corrections, ensuring accurate meaning, refining tone, maintaining terminology consistency, enhancing coherence and readability, and checking for completeness in the translation process.
Grammatical Corrections: Post-editors focus on correcting grammatical errors introduced by the machine translation system. This includes fixing issues related to sentence structure, verb tense, agreement, word order, punctuation, and other grammatical elements.
Mistranslations and Inaccuracies: Post-editors ensure that the translated text accurately reflects the meaning of the source text, rectifying any errors or distortions introduced by the machine translation system.
Tone: Post-editors refine the tone of the translated text to make it suitable for the target audience. This involves adjusting the language register, adapting the content to the desired level of formality or informality, and ensuring consistency throughout the translation.
Terminology Consistency: Post-editors verify and correct terminology used in the translation to ensure it aligns with specified glossaries, style guides, or client preferences.
Coherence and Readability: Post-editors enhance the coherence and readability of the translated text. They rephrase or restructure sentences as needed to improve flow, clarity, and overall comprehension without altering the intended meaning.
Completeness: Post-editors check whether the entirety of the source text has been translated by the engine and flag if anything is missing. Please note that translating the missing elements is not in the scope of MTPE!
What MTPE Doesn’t Cover
Understanding the boundaries of MTPE is crucial for effective collaboration. MTPE projects generally avoid major rewrites, extensive cultural adaptation, additional research, and creative modifications beyond the initial machine translation output.
Major Rewrites: MTPE generally does not involve major or complete rewriting of the translated text. The goal is to work with the existing machine-generated output and improve it rather than starting from scratch, therefore the final text may contain traces of the machine style widely associated with MT translations The rule of thumb is that any text requiring more than 70% rewriting should be submitted for a retranslation process rather than MTPE.
Extensive Localization: MTPE primarily addresses linguistic aspects, so extensive cultural adaptation, such as replacing quotations from books, movies, songs with analogical ones from the target culture, or localizing jokes, puns, and wordplays to reflect local culture, is considered out of scope.
Additional Research and Subject Matter Expertise: Extensive research beyond provided reference materials is considered out of scope for post-editors. While they may use external resources for clarification, the focus is on refining the existing translation rather than conducting in-depth research. Post-editors are not expected to be experts on the subject matter of the text.
Creative Adaptations: Post-editors typically do not engage in creative adaptations or substantial content modifications that go beyond the initial machine translation output. The aim is to retain the meaning of the source text while making necessary improvements.
Quality of the Source Text: While spotting and reporting source text issues frequently occur during human localization, these issues are out of scope for MTPE.
Formatting and Layout: Tasks related to formatting, layout adjustments, or design elements are usually not part of the MTPE process.
The expected outcome of MTPE projects
The desired outcome of MTPE is, therefore, inherently different from that of the editing of human translation. The text that undergoes MTPE is expected to be free of grammatical, punctuation, register, or major style mistakes - coherent, logical, and accurate.
Simultaneously, since MTPE does not involve complete rewriting, the resultant text may retain a machine-like feel, occasional unnatural phrasing, coarseness, ambiguity, and variations in style and tone.
Moreover, it may lack certain linguistic nuances, cultural references, or accurately conveyed idiomatic expressions. Specialized vocabulary may also be prone to inaccuracies in this process.
Risks associated with MTPE projects
Risks exist in MTPE projects, encompassing inconsistencies, accuracy, unpredictable timelines and costs, biases, hallucinations, data safety concerns and copyright issues. Delving into these aspects sheds light on the intricacies and potential pitfalls associated with machine translation processes, ensuring a well-informed approach for both those who order the MTPE services and those who provide the services.
Inconsistencies: In its current state, AI is not capable of being internally consistent, even for simple texts. The post-editing process minimizes the risk but may not eradicate the issue altogether- this heavily depends on the expected turnaround time, budget, and quality of the machine translation.
Accuracy: AI engines do not understand the text they “translate” often leading to a loss of meaning, misinterpretation, or even misleading results. Eradicating those might be very time- and energy-consuming, therefore the final results might be subpar for projects with tight deadlines or considerable budget restrictions.
Unpredictable Timelines and Costs: The quality of machine translation is often very uneven throughout the text: drops and surges in quality are often experienced. This leads to limitations in the initial assessment of the amount of work the project may require and may eventually impact timelines and budget in a considerable way
Bias: Translations provided by AI may reflect and sometimes amplify bias present in the training data, leading to the inclusion of unfair, untrue, or offensive information not present in the source material. If incorporated into the final output, it might present the content in an unintended, unfavorable way, lessen or distort its impact, or lead to issues with the target audience.
Hallucinations: AI hallucinations refer to instances where a language model, such as ChatGPT, generates output that sounds plausible but is, in fact, incorrect, fictional, or disconnected from reality. Careful review and fact-checking are crucial. Since extensive research is not within the typical scope of MTPE, AI hallucinations may find their way into the target, rendering it unusable, misleading, or potentially dangerous.
Data Safety & Risk of Leaking: The confidentiality policies of various AI models are shaky at best, especially if one uses the free versions of available software. Most use input materials and translation results to train their models, potentially feeding confidential content back to the internet without control. In such circumstances, service providers cannot fully protect the confidentiality of leaked content from the AI translation process.
Copyrights: Large language models are not creative; they reuse text found online. This poses a high risk of incorporating copyrighted material into the translation process, exposing users to legal risks. MTPE process is not designed to check whether the original machine output used the copyrighted text, and therefore will not eliminate legal risks related to its use.
While choosing the most appropriate way to localize your content, it is crucial to balance out the gains and losses related to expected quality, timeline, and costs. The misalignment between expectations and the reality of MTPE services may result in delays, disappointments, or a loss of profit on a specific project or, in more severe cases, seriously affect business relationships and customer trust. Successfully navigating the complexities of MTPE requires a clear understanding of its limitations as posed against the expected gains; open communication and effective expectation management between service providers and clients will lead you to the best results.
Ready to implement your game localization with Machine Translation Post-Editing (MTPE)?
Contact LocalizeDirect
today for a complimentary consultation tailored to your specific needs. Let us guide you toward time-efficient, cost-effective, and quality-driven language solutions.
Editing on human-made translation and machine translation offer unique benefits and drawbacks, which impact the cost, quality, and time of the translation and localization projects.
