Skip to Main Content
Table 18

Adaptation using data from target language, related languages, and unrelated languages. All models are trained using NB on the Web1T corpus. In all settings, the total number of examples used to estimate priors is the same. Averaged results for Russian, Polish, Spanish, and Italian. Best results for each phenomenon are emphasized in bold. In the parentheses, relative improvement of each adapted model with respect to the corresponding native-trained model is shown. All improvements of adapted models vs. native-trained models and improvements of the best adapted models vs. models adapted using unrelated language data are statistically significant (McNemar's test, p < 0.001).

Error type(1) Native-trainedPerformance (F1) (2-4) Adapted models
(2) Target(3) Related(4) Unrel.
Article 30.70 32.90 (7.17%) 33.11 (7.85%) 31.77 (3.49%) 
Prep. 22.40 26.01 (16.12%) 25.43 (13.53%) 23.62 (5.45%) 
Verb agr. 34.67 38.76 (11.80%) 37.56 (8.34%) 36.07 (4.04%) 
Error type(1) Native-trainedPerformance (F1) (2-4) Adapted models
(2) Target(3) Related(4) Unrel.
Article 30.70 32.90 (7.17%) 33.11 (7.85%) 31.77 (3.49%) 
Prep. 22.40 26.01 (16.12%) 25.43 (13.53%) 23.62 (5.45%) 
Verb agr. 34.67 38.76 (11.80%) 37.56 (8.34%) 36.07 (4.04%) 
Close Modal

or Create an Account

Close Modal
Close Modal