Model statistics

Test splitMS (multi-signer) - test set includes signers also seen in training (known to the model).SI (signer-independent) - test set contains only signers unseen in training (generalization test).

Uni-Sign (ensemble)

Transfer sourceSplitBLEU-1BLEU-2BLEU-3BLEU-4ROUGE-LBLEURTchrFBERTScore
OpenASL + How2Sign (6 models)MS34.6220.6914.1710.3624.4344.3727.8989.24
OpenASL + How2Sign (2 models)SI30.0015.659.856.6818.6539.1923.4188.35

What the ensembles are made of

MS: six Uni-Sign models chosen one by one on the validation set. They are the OpenASL and How2Sign transfer models, each with two random seeds, a model trained on OpenASL, then How2Sign, then PJM, and a model trained on a mix of the references and translations by other models that had not seen those clips.

SI: two Uni-Sign models, with OpenASL and How2Sign transfer.

An ensemble averages the models' predictions at every next word, so it is not a single-model result, and it costs as much computation as it has models.

Uni-Sign

Transfer sourceSplitBLEU-1BLEU-2BLEU-3BLEU-4ROUGE-LBLEURTchrFBERTScore
OpenASLMS32.6118.4112.108.5321.7242.4326.0488.91
How2SignMS32.4218.3312.118.6121.9041.9825.8688.86
CSL-NewsMS26.4212.697.705.1616.3536.3620.5687.79
none (from scratch)MS18.585.772.961.828.4830.0413.3686.14
OpenASLSI29.2814.939.186.0917.6238.1622.8288.18
How2SignSI28.7814.438.865.9117.3737.8322.3688.14
CSL-NewsSI21.987.894.002.4111.3631.3116.8087.01
none (from scratch)SI15.293.481.520.856.1130.0813.0785.95

SpaMo

Transfer sourceSplitBLEU-1BLEU-2BLEU-3BLEU-4ROUGE-LBLEURTchrFBERTScore
PHOENIXMS28.8713.768.275.5419.6635.1221.0687.79
none (PJM only)MS18.556.603.402.0312.3730.2515.3286.13
PHOENIXSI21.977.463.632.0912.4228.8416.0286.63
none (PJM only)SI17.654.531.860.939.2827.9713.1485.75

SpaMo (mod.)

Transfer sourceSplitBLEU-1BLEU-2BLEU-3BLEU-4ROUGE-LBLEURTchrFBERTScore
PHOENIXMS27.9013.147.765.0618.8634.3020.9487.68
none (PJM only)MS18.456.042.851.5611.9230.4214.5086.02
PHOENIXSI21.977.393.592.0612.2627.8616.1786.53
none (PJM only)SI18.785.232.271.1710.1227.6714.1486.03

Bold marks the best value in each column within its split, MS or SI.