Apppublic

oddadmix/jisr-aligner-demo

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

2likes

App README

Jisr aligners

Two tabs, two models. Sentences takes an Arabic document and its English translation and returns the sentence links with the posterior confidence the DP assigned them - including the deletions, which is where aligners quietly fail. Words takes a single sentence pair and links it token by token, drawing the whole similarity matrix so the near misses are visible.

Models: oddadmix/Jisr-Align-29M and oddadmix/Jisr-WordAlign-50M, bundled under model/ and model-word/. Gold: oddadmix/jisr-align-gold.

del_cost is a slider because it is a property of your corpus, not of the model: it is the price of a deletion, so it decides whether a 1-2 is answered by merging or by inventing a dropped sentence. Machine-translated text that rarely omits anything wants ~0.70; human translation that does omit wants ~0.45. The gap between those two settings is worth 0.06-0.17 F1 depending on the corpus.

The two tabs are deliberately not one pipeline. The word model is measured on gold sentence pairs, so feeding it the sentence aligner's output would compound the two error rates and neither published number would describe what you were looking at.