Linguistic representations in multi-task neural networks for ellipsis resolution

Ola  Rønning; Daniel  Hardt; Anders Søgaard

Linguistic representations in multi-task neural networks for ellipsis resolution

Ola Rønning, Daniel Hardt, Anders Søgaard

Abstract

Sluicing resolution is the task of identifyingthe antecedent to a question ellipsis. Antecedentsare often sentential constituents, andprevious work has therefore relied on syntacticparsing, together with complex linguisticfeatures. A recent model instead used partialparsing as an auxiliary task in sequential neuralnetwork architectures to inject syntactic information.We explore the linguistic informationbeing brought to bear by such networks,both by defining subsets of the data exhibitingrelevant linguistic characteristics, and byexamining the internal representations of thenetwork. Both perspectives provide evidencefor substantial linguistic knowledge being deployedby the neural networks.

Original language	English
Title of host publication	Proceedings of the 2018 EMNLP Workshop BlackboxNLP : Analyzing and Interpreting Neural Networks for NLP
Publisher	Association for Computational Linguistics
Publication date	2018
Pages	66–73
Publication status	Published - 2018
Event	2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP - Brussels, Belgium Duration: 1 Nov 2018 → 1 Nov 2018

Workshop

Workshop	2018 EMNLP Workshop BlackboxNLP
Country/Territory	Belgium
City	Brussels
Period	01/11/2018 → 01/11/2018

Cite this

Linguistic representations in multi-task neural networks for ellipsis resolution. / Rønning, Ola ; Hardt, Daniel ; Søgaard, Anders.

Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP. Association for Computational Linguistics, 2018. p. 66–73.

Research output: Chapter in Book/Report/Conference proceeding › Article in proceedings › Research › peer-review

@inproceedings{8a94915079df42afa8747ef2fbd942f5,

title = "Linguistic representations in multi-task neural networks for ellipsis resolution",

abstract = "Sluicing resolution is the task of identifyingthe antecedent to a question ellipsis. Antecedentsare often sentential constituents, andprevious work has therefore relied on syntacticparsing, together with complex linguisticfeatures. A recent model instead used partialparsing as an auxiliary task in sequential neuralnetwork architectures to inject syntactic information.We explore the linguistic informationbeing brought to bear by such networks,both by defining subsets of the data exhibitingrelevant linguistic characteristics, and byexamining the internal representations of thenetwork. Both perspectives provide evidencefor substantial linguistic knowledge being deployedby the neural networks.",

author = "Ola R{\o}nning and Daniel Hardt and Anders S{\o}gaard",

year = "2018",

language = "English",

pages = "66–73",

booktitle = "Proceedings of the 2018 EMNLP Workshop BlackboxNLP",

publisher = "Association for Computational Linguistics",

note = "2018 EMNLP Workshop BlackboxNLP ; Conference date: 01-11-2018 Through 01-11-2018",

}

TY - GEN

T1 - Linguistic representations in multi-task neural networks for ellipsis resolution

AU - Rønning, Ola

AU - Hardt, Daniel

AU - Søgaard, Anders

PY - 2018

Y1 - 2018

N2 - Sluicing resolution is the task of identifyingthe antecedent to a question ellipsis. Antecedentsare often sentential constituents, andprevious work has therefore relied on syntacticparsing, together with complex linguisticfeatures. A recent model instead used partialparsing as an auxiliary task in sequential neuralnetwork architectures to inject syntactic information.We explore the linguistic informationbeing brought to bear by such networks,both by defining subsets of the data exhibitingrelevant linguistic characteristics, and byexamining the internal representations of thenetwork. Both perspectives provide evidencefor substantial linguistic knowledge being deployedby the neural networks.

AB - Sluicing resolution is the task of identifyingthe antecedent to a question ellipsis. Antecedentsare often sentential constituents, andprevious work has therefore relied on syntacticparsing, together with complex linguisticfeatures. A recent model instead used partialparsing as an auxiliary task in sequential neuralnetwork architectures to inject syntactic information.We explore the linguistic informationbeing brought to bear by such networks,both by defining subsets of the data exhibitingrelevant linguistic characteristics, and byexamining the internal representations of thenetwork. Both perspectives provide evidencefor substantial linguistic knowledge being deployedby the neural networks.

M3 - Article in proceedings

SP - 66

EP - 73

BT - Proceedings of the 2018 EMNLP Workshop BlackboxNLP

PB - Association for Computational Linguistics

T2 - 2018 EMNLP Workshop BlackboxNLP

Y2 - 1 November 2018 through 1 November 2018

ER -