Skip to main content
. Author manuscript; available in PMC: 2019 Dec 4.
Published in final edited form as: Proc Conf Empir Methods Nat Lang Process. 2019 Nov;2019:4240–4250. doi: 10.18653/v1/D19-1434

Table 1:

Dev accuracy results for MT-DNN model with different training set sampling strategies.

Strategy % of Training Data
0.1% 1% 10%
Random (reported) 82.1 85.2 88.4
Random (small batch) 81.79 84.90 88.32
Lower-bound 43.68 41.56 39.89
Upper-bound 81.62 80.46 79.06
AVI 82.44 85.44 86.73
AVO 43.60 42.05 40.81