Recent Semantic Role Labeling (SRL) models have pushed the 
state-of-the-art by using end-to-end supervised neural sequence labeling 
models. The trade-off of this approach is the need for a considerable 
amount of annotated training data, which is currently a problem for 
languages that do not have big SRL datasets. We propose a multilingual 
Encoder-Decoder model that takes advantage of rich existing SRL 
annotations in English and learns to simultaneously translate and 
generate sentences with SRL annotations on a target lower-resource 
language. We achieve this via three steps: i) benchmark our approach on 
well-known SRL datasets ii) train a multilingual model that makes use of 
data in different languages to improve labeling and iii) train a 
cross-lingual model and use it to generate more SRL labeled data on the 
target side. We measure the improvement of SRL models on lower-resource 
target languages (we take the specific cases of French and German) by 
augmenting the training data and also perform a manual evaluation on the 
extended dataset to analyze the quality and precision of the newly 
generated SRL data.