TY - GEN
T1 - Exploring the syntactic abilities of RNNs with multi-task learning
AU - Enguehard, Émile
AU - Goldberg, Yoav
AU - Linzen, Tal
N1 - Publisher Copyright:
© 2017 Association for Computational Linguistics.
PY - 2017
Y1 - 2017
N2 - Recent work has explored the syntactic abilities of RNNs using the subject-verb agreement task, which diagnoses sensitivity to sentence structure. RNNs performed this task well in common cases, but faltered in complex sentences (Linzen et al., 2016). We test whether these errors are due to inherent limitations of the architecture or to the relatively indirect supervision provided by most agreement dependencies in a corpus. We trained a single RNN to perform both the agreement task and an additional task, either CCG supertagging or language modeling. Multitask training led to significantly lower error rates, in particular on complex sentences, suggesting that RNNs have the ability to evolve more sophisticated syntactic representations than shown before. We also show that easily available agreement training data can improve performance on other syntactic tasks, in particular when only a limited amount of training data is available for those tasks. The multi-task paradigm can also be leveraged to inject grammatical knowledge into language models.
AB - Recent work has explored the syntactic abilities of RNNs using the subject-verb agreement task, which diagnoses sensitivity to sentence structure. RNNs performed this task well in common cases, but faltered in complex sentences (Linzen et al., 2016). We test whether these errors are due to inherent limitations of the architecture or to the relatively indirect supervision provided by most agreement dependencies in a corpus. We trained a single RNN to perform both the agreement task and an additional task, either CCG supertagging or language modeling. Multitask training led to significantly lower error rates, in particular on complex sentences, suggesting that RNNs have the ability to evolve more sophisticated syntactic representations than shown before. We also show that easily available agreement training data can improve performance on other syntactic tasks, in particular when only a limited amount of training data is available for those tasks. The multi-task paradigm can also be leveraged to inject grammatical knowledge into language models.
UR - http://www.scopus.com/inward/record.url?scp=85072967346&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85072967346&partnerID=8YFLogxK
U2 - 10.18653/v1/k17-1003
DO - 10.18653/v1/k17-1003
M3 - Conference contribution
AN - SCOPUS:85072967346
T3 - CoNLL 2017 - 21st Conference on Computational Natural Language Learning, Proceedings
SP - 3
EP - 14
BT - CoNLL 2017 - 21st Conference on Computational Natural Language Learning, Proceedings
PB - Association for Computational Linguistics (ACL)
T2 - 21st Conference on Computational Natural Language Learning, CoNLL 2017
Y2 - 3 August 2017 through 4 August 2017
ER -