2017Q3 Reports: SIGNLL

From Admin Wiki
Jump to: navigation, search

The goals of SIGNLL, ACL's special interest group on natural language learning, are to promote and inform about research on computational modeling of learning in natural language. These are served by (i) the maintenance of an informative and up-to-date website and associated mailing list, and (ii) the organization of annual events (the CoNLL conference and the CoNLL shared tasks), and support of other related activities.

The website, located at URL http://www.signll.org and maintained by Ben

Verhoeven, has recently been updated, and remains an important source of information. This is complemented by an email list for announcements for SIGNLL-related events, managed by Erik Tjong Kim Sang. In addition, SIGNLL maintains the domain www.conll.org for specific information about CoNLL conferences.

As of November 1 2016 the president is Julia Hockenmaier (University of Illinois at Urbana Champaign) and the secretary is Suzanne Stevenson (University of Toronto).

Apart from the officers, SIGNLL also has two consultative committees. The SIGNLL Steering Committee, composed by all past SIGNLL officers: Antal van den Bosch, Claire Cardie, Xavier Carreras, Alexander Clark, Walter Daelemans, Lluis Marquez, Hwee Tou Ng, Joakim Nivre, David Powers, and Dan Roth; and the larger SIGNLL International Advisory Board (see http://www.signll.org/officers for a complete description of SIGNLL officers and boards).

The current membership of SIGNLL is 364 people.

CoNLL 2017

In 2017, the annual meeting of CoNLL is collocated with the ACL annual meeting in Vancouver, British Columbia. It will take place August 3-4, 2017. The conference chairs are Roger Levy (MIT) and Lucia Specia (University of Sheffield).

This year the conference accepted only long papers. It received a record 280 submissions, of which 2 were rejected for formal reasons and 12 were withdrawn by the authors. Of the remaining 271 papers, 50 were chosen to appear in the conference program, with an overall acceptance rate of 18.7%, the lowest ever for the conference. Seven of these were withdrawn after the notification, resulting in 43 papers for the final program: 20 selected for oral presentation, and the remaining 23 for poster presentation plus lightning oral presentation.

The invited speakers are Chris Dyer (DeepMind/CMU) and Naomi Feldman (University of Maryland).

More information can be found at the conference website: http://www.conll.org/2017

Although it is unclear whether the record number of submissions this year was due to the fact that there were relatively few other conferences (no NAACL or COLING), the current organizational structure of the conference (two conference chairs, but no area chairs or publication chairs) is not appropriate anymore, and we will work to recruit a larger organizing committee with at least a few area chairs and a publications chair for next year.

It should also be noted that the reviewing periods of CoNLL overlapped with those of EMNLP and *SEM, leading to a number of double or even triple submissions.

CoNLL 2017 Shared Tasks

CoNLL has a tradition of hosting high profile shared tasks. Due to the high quality of proposals, CoNLL 2017 is for the first time hosting two shared tasks instead of a single shared task. Papers of the shared tasks are collected in two separate companion volumes of the CoNLL 2017 proceedings.

The CoNLL SIGMORPHON 2017 Shared Task: Universal Morphological Reinflection is organized by a committee chaired by Mans Hulden (University of Colorado), and co-sponsored by SIGMORPHON. They received a total of 27 submissions for two different subtasks.

The CoNLL 2017 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies

is organized by a committee chaired by Jan Hajič (Charles University, Prague). They received a total of 32 submissions, and used again the TIRA platform from the University of Weimar, managed by Martin Potthast. TIRA allows participants to upload their systems and evaluate them automatically. One of the main advantages of this approach is that the evaluation sets remain confidential.