Segmentation and TM leverage
Does our segmentation reproduce Acolad's closely enough to keep today's Translation Memory leverage? Tested on 212 Windchill topics in 9 languages.
Spike 3 doneSpike 4 done
Step 3 of 24 · Presenter note
This spike protects the business case. If our segmentation differs from Acolad’s, exact TM matches drop and both AI translation and proofreading work go up.
Leverage is preserved within the proposed tolerance
After three rule iterations, boundary agreement is 97.4% and exact-match leverage is 57.1% against Acolad's 57.8%, a loss of 0.7 percentage points. Thai is the exception and needs Acolad's Thai rule.
97.4%Boundary agreement with AcoladIteration 1: 91.2%
57.1%Exact-match leverageAcolad today: 57.8%
4.92MTranslation units imported38 of 38 languages
99.1%Exact-match agreement with AcoladFuzzy bands 93.4% · lookup 45 ms (p95)
Exact-match leverage by language
Illustrative dataAcolad todayNew platform, iteration 3
DE58.3% (-0.6)
FR60.6% (-0.6)
IT61.5% (-0.5)
ES59.7% (-0.4)
PT-BR55.1% (-0.5)
PL56% (-0.4)
ZH56.2% (-1.3)
AR51% (-0.8)
TH42.1% (-3.1)
Rule iterations
| Iteration | What changed | Boundaries | Leverage | Against Acolad |
|---|---|---|---|---|
| 1 | Acolad SRX rules as delivered | 91.2% | 49.3% | −8.5 pp |
| 2 | Abbreviation exceptions, tag-aware boundaries | 95.8% | 55.0% | −2.8 pp |
| 3 | Colon lists, no-break space normalisation | 97.4% | 57.1% | −0.7 pp |
Each iteration re-runs the full corpus against the existing TM, so every rule change shows its effect on leverage, not only on boundaries.
Where segmentation still differs
| Difference | Example | Boundary differences | Status | Fix |
|---|---|---|---|---|
| Abbreviations (approx., e.g., Fig., max., No.) | Tighten screws approx.| 2 turns | 1,412 | Fixed | Added to SRX exceptions for all languages |
| Inline <uicontrol> / <xref> at sentence end | Press <uicontrol>OK</uicontrol>.| Then… | 388 | Fixed | Tag-aware boundary rule |
| Colon-introduced lists | Check the following:| door gap | 205 | Fixed | Keep list lead-in with first item, as Acolad does |
| No-break space in units | 25 Nm vs 25 Nm | 144 | Fixed | Normalise before matching |
| Thai sentence boundaries | No spaces between words | 97 | Open | Acolad Thai rule requested |
| Titles ending with a period | Commissioning. | 61 | Open | Decide: strip or keep |
Example: why one abbreviation matters
English sourceTighten the screws approx. 2 turns, e.g. with a torque wrench (see Fig. 4).
Acolad todayTighten the screws approx. 2 turns, e.g. with a torque wrench (see Fig. 4).
Iteration 1Tighten the screws approx.2 turns, e.g.with a torque wrench (see Fig. 4).
Iteration 3Tighten the screws approx. 2 turns, e.g. with a torque wrench (see Fig. 4).
Split into three units, the sentence no longer matches its TM entry, so it would be translated and reviewed again in all 38 languages.