Turn reference translations into a memory

An alignment run pairs Portuguese documents with their English translations. The type, languages, files, model and confidentiality are set, files are paired by name and by hand, and the run aligns 27 pairs for about five cents. The pairs are reviewed and saved as TMX; a Structural run aligns a web page with no model; and the TMX is imported into a TM that then gives 100% matches.

1. When to align

   You have documents that were translated before, but no TM. TM alignment pairs the sentences of each document with its translation and writes the pairs as a TMX, ready to import into a translation memory.

   

2. Open TM Alignment and click Create New Alignment Run

   Open TM Alignment. Each alignment is a run: Create New Alignment Run starts one, and a saved run opens again where it was left, or at its result.

   

3. Name the run and choose the type

   Name the run and choose the type. Lightweight uses an AI model. Structural is for web pages built alike. External App needs an alignment server, and is greyed out here because none is set up. The notice below says what happens to PDFs. The type is fixed once a file is added.

   

4. Set the languages, add the files, choose the AI model and Confidentiality

   Set the languages by hand: the files don't say what language they are in, and the TMX is labelled with this pair. Add the Portuguese files as sources and the English files as targets. Choose the AI model with Change…. Under Confidentiality you can choose profiles that mask names and numbers before anything is sent. These leaflets are public, so none is chosen, and the screen says the text is sent as written. A batch holds at least 150 Portuguese words.

   

5. Click Auto-pair by Name; pair the rest by hand with Pair Selected

   Auto-pair by Name pairs guia-do-visitante_pt with guia-do-visitante_en: the names agree once the language code is removed. The web pages have different names, so pair them by hand: click one in each list, then Pair Selected. A file goes into one pair only, and an unpaired file is not aligned.

   

6. Click Start Alignment

   Click Start Alignment and confirm. The text is sent in batches, one request per batch: two here. The counters show the pairs aligned, the tokens, and the estimated cost at the AI model's list price, which is billed to your own API key. Cancel Process stops before the next request, and the requests already sent are still charged.

   

7. Read the result: nothing has been reviewed yet

   27 pairs from two documents, for about five cents. Nothing has been reviewed: a sentence paired with the wrong one goes into the TMX as it is, and into every match it gives later. AV does not edit the pairs, so read them before the TMX goes into a TM.

   

8. Click View Results and check the pairs

   View Results shows the pairs as they will be saved. Pair 3 is one Portuguese sentence translated as two English ones; pair 5 is two Portuguese sentences translated as one. Both are right: each pair says the same thing on both sides. A wrong pair would have English that says something else, and the rows after it would usually be off too.

   

9. Click Finish & Save TMX

   Click Finish & Save TMX and choose where to save it. The TMX holds each pair with the name of the document it came from. The run stays in the list as complete; open it again to view the pairs or save another copy.

   

10. A Structural run: no model, nothing sent

    A second run, Structural, on the accessibility page alone. No model and no prompt: the pages are paired on this computer and nothing is sent. Pair the two pages by hand, as before.

    

11. Check the Structural result

    Structural pairs the 10 numbered paragraphs, which match one for one, and leaves out the heading and the footer. It needs pages built alike. When it finds nothing to pair it says so and suggests Lightweight.

    

12. In a project, open TMs, tick the TM and click Import TMX

    The saved file is an ordinary TMX. Open a project with the same language pair, go to TMs, tick the TM and click Import TMX. All 27 pairs go into the TM.

    

13. The first matches at 100%

    The museum's next leaflet repeats two sentences from the visitor guide word for word, and the TM offers both at 100%. A sentence that is only close to an old one gets no match here.

    

14. Recorded on 20/10/2026

    Video recorded on 5 November 2026 with AV 0.2.11; the alignment and its costs on 20 October 2026. Costs are estimates at the list prices in AV's model settings on that date.