Relevance of Linguistic Structure in Neural NLP


  1. Chris Dyer
  2. Emily Bender
  3. Jason Eisner
  4. Mark Johnson


Call for long and short papers!

Long papers may consist of up to eight (8) pages of content, plus unlimited references; final versions of long papers will be given one additional page of content (up to 9 pages) so that reviewers’ comments can be taken into account.

Short papers may consist of up to four (4) pages of content, plus unlimited references. Upon acceptance, short papers will be given five (5) content pages in the proceedings.


There is a long standing tradition in NLP focusing on fundamental language modeling tasks such as morphological analysis, POS tagging, parsing, WSD or semantic parsing. In the context of end-user NLP tasks, these have played the role of enabling technologies, providing a layer of representation upon which more complex tasks can be built.

However, in recent years we have witnessed a number of success stories for tasks ranging from information extraction or text comprehension to machine translation, for which the use of embeddings and neural networks has driven state of the art results to new levels. More importantly, these are often end-to-end architectures trained on large amounts of data and making little or no use of a linguistically-informed language representation layer. For example, the modeling of word senses and word sense disambiguation are implicit in the functional composition of word embeddings. Other topics such as linear sentence processing versus syntactic parses or frequency-based word segmentation versus morphological analysis are still up for debate.

This workshop focuses on the role of linguistic structures in the neural network era. We aim to gauge their significance in building better, more generalizable NLP systems. We would like to address the following questions:

  1. Is linguistic information useful for neural network architectures: can it improve state of the art neural architectures, and how should it be used? Does it help in building models that transfer better to new domains, new languages, new tasks, or for other limited annotated data scenarios?
  2. Are there any better implicit representations that neural networks can extract, whether similar or not to linguistic structures, that can be transferred or shared across tasks and, hence, serve as core language representation layers?