Azure Translator
An Azure service to easily conduct machine translation with a simple REST API call.
461 questions
How to Resolve Random Placement of Placeholders in Microsoft Custom Translator Output?
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(224.00000000000003, 60%, 31%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3ETA%3C/text%3E%3C/svg%3E)
Tapia, Alejandra
0
Reputation points
I am working with two language pairs (en-> fr), (en-> es), and the custom models are working fine, but they contain many placeholders and those are just being placed randomly in the target text.
This, of course, affects the translation output since those placeholders are sentence subjects/objects/adjectives, so verbs end up also being conjugated randomly due to a "lack of subject" or, as shown in the example, prepositions are missing and the sentence changes in meaning.
Are there settings within the Custom Translator that can help fix this?
Could this be a problem related to the CAT tool that's being used?
My training data is ~10,000 segments that do include placeholders.
Example:
| en source | es MT output | expected output |
|---|---|---|
| These changes mostly affect the global navigation bar, displayed at the top of all {1} pages. | Estos cambios afectan principalmente a la barra de navegación global, que se muestra en la parte superior de todas las {1} páginas. | Estos cambios afectan principalmente a la barra de navegación global, que se muestra en la parte superior de todas las páginas de {1}. |
{count} votes
-
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(6.4, 20%, 10%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3EKS%3C/text%3E%3C/svg%3E)
kothapally Snigdha 2,185 Reputation points Microsoft External Staff2025-03-20T22:57:33.5066667+00:00 It seems like you're encountering challenges with placeholders in your translations, which can indeed affect the quality of the output. Let me address your concerns step by step:
- Custom Translator allows for some control over placeholders. You can define placeholder tags in your training data to ensure they are treated appropriately during translation. This involves marking placeholders explicitly in your data so that the model recognizes them as non-translatable elements. Check the documentation for guidance on how to format and tag placeholders effectively.
- Additionally, you might want to experiment with the "Preprocessing" and "Postprocessing" options in the Custom Translator. These settings can help manage placeholders and ensure they are correctly positioned in the output.
- CAT tools (Computer-Assisted Translation tools) can sometimes mishandle placeholders if they are not properly configured. Ensure that your CAT tool is set up to recognize and preserve placeholders during the translation process. Many CAT tools allow you to define rules or patterns for placeholders, which can help maintain their integrity in the target text.
- If your CAT tool is not handling placeholders well, it might be worth exploring alternative tools or plugins that offer better support for placeholder management.
- Since your training data includes placeholders, it's crucial to ensure they are consistently formatted and tagged. Inconsistent or ambiguous placeholder usage in the training data can confuse the model and lead to errors in the output.
- Consider augmenting your training data with additional examples that demonstrate correct placeholder usage. This can help the model learn how to handle placeholders more effectively.
- In your example, the expected output places the placeholder {1} differently than the MT output. This suggests that the model or tool is not correctly interpreting the placeholder's context. Ensuring proper tagging and alignment in your training data, as well as configuring your CAT tool appropriately, should help address this issue.
I hope these helps you, Thank you!
-
kothapally Snigdha 2,185 Reputation points Microsoft External Staff
2025-03-22T00:20:41.9366667+00:00 just checking back to see if the above response was helpful.
-
kothapally Snigdha 2,185 Reputation points Microsoft External Staff
2025-03-24T20:22:43.7333333+00:00 We haven’t heard from you on the last response and was just checking back to see if you have a resolution yet. In case if you have any resolution, please do share that same with the community as it can be helpful to others. Otherwise, will respond with more details and we will try to help.
-
Tapia, Alejandra 0 Reputation points
2025-03-25T19:31:01.0566667+00:00 Hi @Snigdha Kothapally (Quadrant Technologies) :
The team is cleaning data to try the training solution with more examples. In the meantime, I wanted to ask: how can this situation have to do with the CAT tool? Since the text is being translated/processed by the MS custom translator, the vendor told us the issue wasn't coming from their side, but I'd like to understand if it's possible that the translation software is actually interfering with the custom model's processing.
-
kothapally Snigdha 2,185 Reputation points Microsoft External Staff
2025-03-25T22:35:09.95+00:00 - The random placement of placeholders in the output of Microsoft Custom Translator could potentially be influenced by how the translation software processes the text. If the text being translated contains embedded tags or placeholders, the alignment and processing of these elements can affect their placement in the final output.
- Custom Translator aligns sentences based on document markup and attempts to match corresponding sentences in the source and target languages. If the CAT tool or the source documents have inconsistencies in how placeholders are used or if they are not correctly aligned, this could lead to unexpected results in the translation output.
- It's essential to ensure that the source documents are clean and properly formatted, as any discrepancies in the structure could interfere with the model's ability to process the content accurately. Additionally, reviewing the training data for any issues related to placeholder usage might help in identifying the root cause of the problem.
I hope these helps you. Thankyou!
-
kothapally Snigdha 2,185 Reputation points Microsoft External Staff
2025-03-27T20:34:11.0033333+00:00 We haven’t heard from you on the last response and was just checking back to see if you have a resolution yet. In case if you have any resolution, please do share that same with the community as it can be helpful to others. Otherwise, will respond with more details and we will try to help.
Sign in to comment
Sign in to answer